Home

Intelligent Alarms in Anesthesia: An implementation

1. 3 4 Pressure in the inspiratory limb 3 5 Flow in the expiratory hose Chapter 4 Feature analysis 4 1 Introduction uo iste T 4 2 Normal curves versus malfunctions 4 3 Detection rules Chapter 5 An introduction in expert systems 5 1 What is an expert system En 5 1 1 Knowledge base 5 1 2 Inference engine i 5 2 SIMPLEXYS an expert system building 5 2 1 The SIMPLEXYS rule compiler 5 2 2 A SIMPLEXYS program 5 3 Implementation in SIMPLEXYS Intelligent Alarms in Anesthesia e gt 9 o a e s o o o page ii vi vii Gh Ut ui 4 yr aA KA Mo NIMM M Ka Kad Ka Ka HIE Ka eee d OY U GJ NN f KA KA KA NN KA KA Ln G 0 0 WW WN MMM KM PA GG W WQ GSH iii Chapter 6 Program considerations s e e e e e gt o 33 6 1 Introduction a uu vet fo aor ieee RT m Tera feels 33 6 1 1 Signal analysis E E a DV Tex se RU fa 33 6 1 2 Feature analysis s jer 34 6 1 3 Rule evaluation a 34 6 2 MultiDos Pl s 4 e X ox e van r Lua Lid tas s 35 6 3 Intertask communications s s e s 35 6 3 1 Information block s s e s s s o 36 6 4 Implementation ge 37 6 4 1 Flow measurement ur xD 38 Chapter 7 Conclusions an Recommendations 40 7 1 Concl sionS ss 4 03 09 1909 ox we er y 40 7 2 Recommendations 2 0x UE s e el oue oe A o 40
2. 4 3 Detection rules From the measurements described in the previous paragraph the attributes up down or unchanged can be added to each feature in case of a malfunction A description of every malfunction can be made in the form of rules that describe the status of the feature set The detection rules were derived from the differences between the values of the normal and the disturbed features A straight forward translation of the observations for rule BS CO2 ABSORBER would be FLOW normal and PRESSURE normal and CO2 INS up and CO2 DO STR normal Now for all rules we take the following three steps l delete all features that are normal normal is the default In this example this gives the following rule CO2 INS up 2 delete features that are superfluous especially if they are unreliable or noisy In this example there are none of those 3 add enough features to make the rule unique i e prevent the rule from being true if other problems arise In this example CO2 INS up does occur with other problems eg INCOMP INS VALVE and INCOMP EXP VALVE see the list above This step is the most difficult it requires a thorough expert level understanding of the problem This step may reintroduce features that are normal or features that are not up or not down In the example the following rule is obtained CO2 INS up and FLW EX VOL unchanged and not FLW MIN down After these steps no two rules should be equivalent nor shoul
3. 90 CO2 UP STR 100 70 OBST ET TUBE there is an obstruction in the endotracheal tube flow FLW EX VOL 36 33 FLW T CONST 1 2 5 FLW MAX 32 15 pressure PRS MAX 13 28 PRS SLOPE 6 gt 13 PRS T CONST 1 3 noisy CO CO2 UP STR 100 gt 65 OBST_INSP_HOSE there is an obstruction in the inspiratory hose flow FLW_EX_VOL 36 33 pressure PRS MAX 13 gt 29 normal means that all the features of this signal are normal Ek It takes a few breaths to reach this value Intelligent Alarms in Anesthesia 21 PRS SLOPE 6 13 CO normal OBST EXP HOSE there is an obstruction in the expiratory hose flow FLW T CONST 1 gt 2 5 FLW_MAX 32 15 pressure PRS T CONST 1 1 7 CO normal DISC Y PIECE there is a disconnect at the Y piece flow FLW EX VOL 36 7 FLW MAX 32 20 pressure gignal flat CO signal flat DISC INSP HOSE there is a disconnect of the inspiratory hose flow FLW EX VOL 36 8 FLW MAX 32 gt 20 pressure Signal flat co signal flat DISC EXP HOSE there is a disconnect of the expiratory hose flow FLW_EX_VOL 36 7 FLW MAX 32 20 pressure Signal flat COs Signal flat DISC FGF there is a disconnect of the Fresh Gas Flow flow FLW EX VOL 36 20 FLW MAX 32 22 pressure PRS MAX 13 gt 4 PRS T CONST 1 0 5 CO CO2 DO STR 118 180 CO2 UP STR 100 80 DISC VENT HOSE ther
4. LITERATURE s We cer oa a s RT Sce B e ie c LU DW EC 42 APPENDIX A Programming tools s s a s so eee ee ee 44 APPENDIX B Manual utility routines e a a o 45 APPENDIX C OMHEDA 5410 VOLUME MONITOR s s a e o 55 Intetligent Alarms in Anesthesia iv LIST OF FIGURES Figure Figure Figure Figure Inspiratory valve Normal CO signal 1 1 2 3 Figure 3 Figure 3 Figure 3 Normal flow signal Figure 4 CO features 4 4 4 4 4 5 Figure ae co features Figure Pressure features Pressure features Flow features Figure Figure Figure Flow features The building of an PAU RP WM Ka NA a M ra Figure Intelligent Alarms in Anesthesia expert Block diagram anesthesia system Circle breathing circuit Simple electrical lung model Normal pressure signal s system page 15 16 17 18 24 24 24 24 24 24 28 LIST OF TABLES Table I CO signal Table II Pressure signal Table III Flow signal Table IV Extracted features Inteltigent Alarms in Anesthesia page 15 17 18 19 INTRODUCTION This project was done as part of the fulfillment of the requirements for a M Sc degree in Electrical Engineering at the Eindhoven University of Technology Division of Medical Electrical Engineering The Netherlands The research was performed at the University of Florida College of Medicine Department of Anesthesiolo
5. 89 E 229 ISBN 90 6144 229 X Met lit opg reg SISO 608 1 UDC 616 089 5 NUGI 742 Trefw anesthesie pati ntbewaking SUMMARY Abnormal and potentially dangerous fault conditions in the anesthesia breathing circuit include leaks obstructions disconnects incompetent valves and CO absorber malfunction Because the current alarms of monitoring devices are not specific enough there is a need for an intelligent alarms system that can determine the integrity of the breathing circuit Three signals were measured at three different places CO partial pressure at the Y piece airway pressure at the patient side of the inspiratory valve and airway flow at the patient side of the expiratory valve From these measurements features were extracted for each signal and the feature changes were analyzed From these analysis a rule base was designed which was implemented in an expert system by using the SIMPLEXYS expert systems language Real time signal analysis and feature extraction were implemented in a multitasking environment Initial tests were performed It was found that an implementation of the total system was possible on an IBM AT with the multitasking environment MultiDos Plus The intelligent alarms system was able to distinguish the following malfunctions incompetent inspiratory valve incompetent expiratory valve exhausted CO absorber obstruction at the Y piece obstruction at the inspiratory hose and obstruction at the ex
6. CO pressure and volume monitors are now standard and widely available In the Ohmeda Modulus II the Ohmeda 5200 CO monitor the Ohmeda 5500 airway pressure monitor and the 5410 volume monitor are used Other signals like O percentage oxygen saturation blood pressure and anesthetic agent percentage are usually measured by additional monitors In a system where flow and CO monitoring are the major features we choose to use CO pressure and flow signals 2 3 2 Previous work Previous work has been done in this field by Rob Bastings BAS87 and Jan van der Aa AA87 They measured CO pressure and flow signals at the Y piece As result of these measurement they could detect 5 clusters of malfunctions cluster 1 Endotracheal tube leak Leak in expiratory hose Ventilator tube leak cluster 2 Leak in inspiratory hose Partial disconnect of the fresh gas flow Intelligent Alarms in Anesthesia 9 cluster 3 Incompetent inspiratory valve cluster 4 Incompetent expiratory valve Exhausted CO absorber cluster 5 Increased airway resistance obstructions It was not possible to distinguish between malfunctions within a cluster The advantage of an integrated sensor at the Y piece that they used is that there is only one sensor that has to be placed The disadvantage is that malfunctions far from the sensors are hard to pick up and because all sensors are located in one place the detection is not optimal Ideally we would
7. THEN DO message Incompetent expiratory valve FLW MIN DOWN There is reverse flow BTEST flw min DN PROCESS ON exit FROM running TO 00 The code in the DECLS part is Pascal code with the procedures and Pascal variables definition 09 The Pascal code in the INITG part is executed only when the expert system program is started up For example if a serial port has to be initialized 13 The Pascal code in the INITR part is executed when a new run of the expert system is started 17 The RULES part contains the SIMPLEXYS rules with the knowledge If tbe exit rule becomes true if a key is hit the program stops looping In the running rule the goal of the expert system is defined The inference engine will try to evaluate the goal rule Intelligent Alarms in Anesthesia 31 line 19 line 22 line 23 line 27 line 27 line 33 line 34 BTEST is a boolean test which returns the evaluation of the following test A STATE rule defines which goals are to be evaluated if that STATE rule s value is true Here the only goal is INCOMP EXP VALVE Rules can have one of four values True TR False FA Possible PO and Undefined UD Initially all the rules are UD When a rule is evaluated it becomes TR FA or PO If PO is the result of an evaluation neither TR nor FA could be assigned to the rule In that case an alternative path of evaluation could possibly be followed to reach a conclusion With the INITIA
8. 6 Chapter 2 Signal measurement versus detection 2 1 Introduction Our goal is to derive a conclusion about the integrity of the breathing circuit from information present in the signals measured in or near the breathing circuit Many things can go wrong with the breathing circuit of an anesthesia machine With the current technology it is difficult to conclude what exactly goes wrong because several different monitors may give partially overlapping and quite unspecific alarms Our first task is to find out which malfunctions should be detected and alarmed on Secondly we have to determine which signals are to be measured in the breathing circuit and where we should measure these signals in order to best detect these malfunctions o W to tec The circle breathing circuit exists of disposable plastic hoses unidirectional valves and a CO absorber Because the hoses have to be easily replaceable leaks and disconnects can occur If the hoses kink an obstruction may result As indicated in 1 6 the soda lime of the CO absorber can get exhausted resulting in inadequate ventilation by CO rebreathing The uni directional valves consists of a disc that elevates when the pressure on one side higher is than the pressure on the other side The disc s movement is limited by a retainer The disc is made of flexible plastic and when it becomes moist humidity of expired gas is high it is possible that the disc gets stuck in the open positi
9. a breath within a certain time eg 10 seconds all the features of that signal will be set to not valid NV and the breath detected flag will be set The advantage of this method is that the information that is available will be analyzed even if some other information is missing The feature analysis is on a breath to breath basis 6 1 3 Rule evaluation The rule evaluation part written in the SIMPLEXYS expert systems language see chapter 5 takes the symbolic information from the feature analysis part evaluates the rules and displays the result of the evaluation alarm messages The rule evaluation is also performed on a breath to breath basis If all three parts have to be incorporated into the same program the problem is that one part has to run 20 times per second and two other parts have to run every breath eg approximately every 6 seconds Either the feature analysis with the rule evaluation has to run in 1 20 second or there has to be a multi tasking operating system where several tasks can run at the same time The computing power and memory of an IBM AT or compatible is sufficient for this task However the operating system mostly Intelligent Alarms in Anesthesia 34 used for these machines MS DOS is a single tasking operating system There are several multi tasking operating systems for the IBM AT eg XENIX a sort of UNIX or MS WINDOWS a graphical multi tasking environment but these programs tend to demand a
10. lot of memory and or a special compiler There is a multi tasking shell on the market that does not have these drawbacks and that accommodates our needs MultiDos Plus by Nanosoft Associates Natick MA 2 MultiDos Plus MultiDos Plus is a multi tasking extension to MS DOS It allows the users to load multiple programs in their computer and have them run concurrently At start up the MultiDos Plus program will replace the interrupt driven system services for screen and keyboard I O disk access and DOS functions with its own routines MultiDos Plus assigns CPU time slices to the various programs Programs with higher priority receive more time slices than programs with lower priority The time slice interval is based on the timer interrupt in the computer and is about 55 milliseconds in the IBM AT Programs which cannot execute for any reason waiting for keyboard input waiting for disk access etc are not assigned time slices A program that runs under MultiDos Plus can be in one of two states foreground or background Only one program at a time can be in the foreground If a program is in the background screen I O is written to an invisible screen There are commands to put a program in the background or foreground MultiDos Plus provides message queues which can be read by any task to make communication data or task control possible eg timing by signal and wait MUL86 6 3 Intertas lt i In our system three tasks h
11. the average flow over the period delta RS 232 IBM AT XT interrupts Because the flow pulses will come at an irregular time interval the best way to count them is a method based on interrupts The IBM machines provide two hardware interrupt channels for communications one for COM1 one for COM2 Four classes of these hardware interrupts are possible 00 change in modem status register 01 transmitter holding register empty 10 data received 11 reception error or break condition received We choose to use the interrupt when data is received Any of the other interrupts could also be used If this interrupt is enabled it gives a hardware interrupt when pin RD 3 on DB25 2 on DB9 goes from low to high Interrupt 0B is given if COM2 is used 0C if COM1 is used The following part of C code should be used to initialize the registers of the serial port COM1 outp Ox3fb 0 clear line control register outp Ox3fc OxOB set RTS DTR user aux input 2 inp Ox3fd read line status register outp 0x3f9 1 enable interrupt on data received In order to use any hardware interrupt the corresponding bit in the interrupt mask register IMR has to be set For the communication interrupt of COM1 this is done by Intelligent Alarms in Anesthesia 55 t in al 21H get the content of IMR and al efH set bit for COM1 out dx 21H send byte iti ice Because we use an interrupt to update our counte
12. the time constants Intelligent Alarms in Anesthesia 41 LITERATURE AA87 BAS87 BLO88 CHE85 coo78 G0087 LIC74 MUL86 COHM83 J Aa J J van der Monitoring the integrity of the circle breathing system during general anesthesia with mechanical ventilation M Sc thesis Department of Electrical Engineering University of Florida Gainesville 1987 Bastings R H A Towards the development of an intelligent alarm system in anesthesia M Sc thesis 1987 Faculty of Electrical Engineering Eindhoven University of Technology 1989 EUT Report 89 E 227 Blom J A The SIMPLEXYS expert systems toolbox Preliminary version of a Ph D thesis chapter 7 to be submitted to the Eindhoven University of Technology in 1990 Cheney E W and D R Kincaid Numerical mathematics and computing 2nd ed Belmont Cal Wadsworth 1985 Cooper J B and R S Newbower C D Long B McPeek Preventable anesthesia mishaps A study of human factors Anesthesiology Vol 49 1978 p 399 406 Good M L and S Lampotang G L Gibby J 5 Gravenstein Training in anesthesiology Critical event simulation Scientific Exhibit Annual Meeting of the American Society of Anesthesiologists Atlanta Ga 10 14 Oct 1987 Lichtiger M and F Moya eds Introduction to the practice of anesthesia New York Harper amp Row 1974 MultiDos a Multi tasking program for PC DOS Program manual Nanosoft Associate
13. 1989 ISBN 90 6144 222 J fwiak L FULL DECOMPOSITION OF SEQUENTIAL MACHINES EUT Report 89 E 223 1989 ISBN 90 6144 223 0 Book of abstracts of the first Benelux Japan Workshop on Information and Communication Theory Eindhoven The Netherlands 3 5 September 1989 Ed by Han Vinck EUT Report 89 E 224 1989 ISBN 90 6144 224 9 Hoeijmakers M J K POSSIBILITY TO INCORPORATE SATURATION IN THE SIMPLE GLOBAL MODEL OF A SYNCHRONOUS MACHINE WITH RECTIFIER EUT Report 89 E 225 1989 ISBN 90 6144 225 7 Dahiya R P and E M van Veldhuizen W R Rutgers L H Th Rietjens XPERIMENTS ON INITIAL BEHAVIOUR OF CORONA TENERATED WITH EL PULSES SUPERIMPOSED ON DC BIAS EUT Report 89 E 226 1989 ISBN 90 6144 226 5 Bastings R H A TOWARD THE DEVELOPMENT OF AN INTELLIGENT ALARM SYSTEM IN ANESTHESIA EUT Report 89 E 227 1989 ISBN 90 6144 227 3 Hekker J J COMPUTER ANIMATED GRAPHICS AS A TEACHING TOOL FOR THE ANESTHESIA MACHINE SIMULATOR EUT Report 89 E 228 1989 SBN 90 6144 228 1 Oostrom J H M van INTELLIGENT ALARMS IN ANESTHESIA An implementation EUT Report 89 E 229 1989 ISBN 90 6144 229 X
14. 4 1 Flow measurement Ohmeda provided three monitors to measure the necessary signals in the breathing circuit the Ohmeda 5200 CO monitor the Ohmeda 5500 airway pressure monitor and the Ohmeda 5410 volume monitor The CO and the pressure monitors have an analog output at the back These signals are easy to sample with an A D converter board in the PC The volume monitor does not provide an analog flow signal instead it provides a pulse every time 3 ml gas has passed through the sensor A signal that can be either high or low determines the direction of the flow A routine that counts Intelligent Alarms in Anesthesia 38 the pulses was written An interrupt is generated for every pulse and the interrupt service routine counts the pulses and determines the flow direction Appendix C describes this routine Intelligent Alarms in Anesthesia 39 Chapter 7 Conclusions an Recommendations 2 1 Conclusions It was possible to build a working prototype of the intelligent alarms expert system on an IBM AT in the multitasking environment of MultiDos Plus Tests were performed on a second data set that was obtained from the anesthesia simulator this is another data set than the development set The system was twice as fast compared to a run in real time The following malfunctions could be identified Incompetent inspiratory valve Incompetent expiratory valve Exhausted CO absorber Obstruction at the Y piece Obstruction in the in
15. 90 6144 214 1 Hoeijmakers M J en J M Vleeshouwers EEN MODEL VAN DE SYNCHRONE MACHINE MET GELIJKRICHTER GESCHIKT VOOR REGELDOELE I NDEN EUT Report 89 E 215 1989 ISBN 90 6144 215 X Pineda de Cyvez J LASER A LAyout Sensitivity ExploreR Report and user s manual EUT Report 89 E 216 1989 ISBN 90 6144 216 8 Duarte J L MINAS An algorithm for systematic state assignment of sequential machines computational aspects and results EUT Report 89 E 217 1989 ISBN 90 6144 217 6 Kamp M M J L van de SOFTWARE SET UP FOR DATA PROCESSING OF DEPOLARIZATION DUE TO RAIN AND ICE CRYSTALS IN THE OLYMPUS PROJECT EUT Report 89 E 218 1989 ISBN 90 6144 218 4 Koster G J P and L Stok FROM NETWORK TO ARTWORK utomatic schematic diagram generation CUT Report 89 E 219 1989 ISBN 90 6144 219 2 Willems F M J CONVERSES FOR WRITE UNIDIRECTIONAL MEMORIES EUT Report 89 E 220 1989 ISBN 90 6144 220 6 Kalasek V K l and W M C van den Heuvel L SWITCH A PC program for computing transient voltages and currents during S off three phase inductances EUT Report 89 E 221 1989 ISBN 90 6144 221 8 Eindhoven University of Technology Research Reports ISSN 0167 9708 Faculty of Electrical Engineering Coden TEUEDE 222 223 224 225 226 227 228 229 J fwiak L THE FULL DECOMPOSITION OF SEQUENTIAL MACHINES WITH THE SEPARATE REALIZATION OF THE NEXT STATE AND OUTPUT FUNCTIONS EUT Report 89 E 222
16. K Y PIECE LEAK PATTERN and PRS SLOPE down and CO2 DO TIME up Intelligent Alarms in Anesthesia 27 Chapter 5 An introduction in expert systems 5 1 What is an expert system Ever since the invention of the computer man tried to let computers think like humans Computers are mainly used to do straight forward things which usually contain a lot of repetition This is mostly non intelligent work In order to let a computer solve problems like humans do that computer program has to be intelligent One approach to make a program intelligent is to provide it with lots of high quality specific knowledge about some problem area These programs are called expert systems The biggest problem with building an expert system is to define what knowledge should be used how to obtain the knowledge and how to implement it Figure 5 1 shows the participants in building an expert system and their relations Domain Toolbuilder E t per Extends and test Expert System Knowledg Building Tool engineer Bullds refines and tests Clerical Staff Figure 5 1 The building of an expert system From Waterman p 8 WAT86 The domain expert is the person who has the knowledge about the particular problem area The knowledge engineer is the person who collects the knowledge and implements it in the expert System The knowledge can be acquired by interviewing the domain Intelligent Alarms in Anesthesia 28 expert An expert system building t
17. LLY keyword a rule can be assigned a initial value other than UD In this example rule INCOMP EXP VALVE needs only the evaluation of rule FLW MIN DOWN THEN DO is used if an action has to be performed when the rule evaluates to true The PROCESS section describes the dynamics of the rule evaluation process when to evaluate which rules The expert system program continuously loops until the exit rule becomes true 5 3 R dd SIMPLEXYS proved to be a useful tool for the implementation of our intelligent alarms system It provides both a fast expert system and an easy interface with a powerful language like Pascal A language like Pascal is needed because interfacing routines which interface with other parts of the system have to be written Intelligent Alarms in Anesthesia 32 Chapter 6 Program considerations 6 1 Introduction Our intelligent alarms system can be divided into the following three functional parts each of which has been described in previous chapters Signal analysis chapter 3 Feature analysis chapter 4 Rule evaluation chapter 5 In this chapter we describe how these parts work together The problems encountered during the implementation and their solutions are described 5 1 1 Signal er We assume that the data is sampled with a frequency of 20 Hz These samples are the input for the signal analysis routines The signal analysis consists of the following parts see chapter 3 phase detect
18. UTE Eindhoven Research Report ISSN 0167 9708 University of Technology Coden TEUEDE N etherlands Faculty of Electrical Engineering Intelligent Alarms in Anesthesia An implementation by J H M van Oostrom EUT Report 89 E 229 ISBN 90 6144 229 X October 1989 Eindhoven University of Technology Research Reports EINDHOVEN UNIVERSITY OF TECHNOLOGY Faculty of Electrical Engineering Eindhoven The Netherlands ISSN 0167 9708 Coden TEUEDE INTELLIGENT ALARMS IN ANESTHESIA An implementation by J H M van Oostrom EUT Report 89 E 229 ISBN 90 6144 229 X Eindhoven October 1989 This report was submitted in partial fulfillment of the requirements for the degree of Master of Electrical Engineering at the Eindhoven University of Technology The Netherlands The work was carried out from March 1988 until December 1988 under responsibility of Professor J E W Beneken Ph D Division of Medical Electrical Engineering Eindhoven University of Technology at the Department of Anesthestology College of Medicine University of Florida Gainesville Florida under supervision of M L Good M D J S Gravenstetn M D and J J van der Aa M E CIP GEGEVENS KONINKLIJKE BIBLIOTHEEK DEN HAAG Oostrom J H M van Intelligent alarms in anesthesia an implementation by J H M van Oostrom Eindhoven Eindhoven University of Technology Faculty of Electrical Engineering Fig tab EUT report ISSN 0167 9708
19. X B Manual utility routines Summary int error str char str Description error pointer to error string to print The funtion error prints the error string on the screen prints a dos error string if possible and aborts the program that calls the function Return Value No return value See Also Example exit on error char filename 80 FILE data if data fopen filename error filename wwe NULL the program is aborted if the file cannot be opened Intelligent Alarms in Anesthesia 45 get features Summary int get features address len queue int address address of message int len number of bytes in message int queue queue number 1 31 Description The function get features gets len bytes for the message queue queue The bytes are stored consecutive starting at address so compiex structures can also be send This function is designed for the small model of the compiler That means that all data is in one segment and that the addresses only consist of the offset An address is a pointer to a variable a If there is no message in the queue the function starts waiting until one arrives Return Value There is no return value See Also send features send control get control multi dos test Example get a message out of queue 5 struct SIGNAL signal frame 3 defined in frame h get features signal fram
20. able to describe the signal as a sequence of segments phases We have to determine which attribute a signal has in which segment and hence which features need to be calculated when 3 2 3 P lidati Parameter validation is done for EXPON and SLOPE In our algorithms if less than ten samples are available for slope calculation the feature value is not valid the calculated slope will not be reliable enough All the other features are always calculated because at this stage the intelligence to determine if a feature is valid or not is not available This will be available in a later stage an expert system 3 3 CO signal The CO signal measured at the Y piece consists of 2 major parts 1 the plateau during expiration 2 the zero plateau during inspiration status attribute feature 1 HORIZ Inspired CO level 2 SLOPE Up slope 3 HORIZ Expired CO level 4 SLOPE Down slope Table I CO signal Inteltigent Alarms in Anesthesia 15 Figure 3 1 Normal CO signal During inspiration gas fresh gas and gas through the CO absorber is forced into the patient This gas normally does n ot contain any CO During expiration the CO level will increase until it reaches a level that is approximately equal to the alveolar CO level The CO signal at the Y piece is described in table I and in figure 3 1 ato A breathing circuit with a lung can be modeled by a resistor capacitance circuit Figure 3 2 Simple ele
21. agnitude ax b y 3 2 The sum of the errors over all m data points is m E ax b y 3 3 k 1 Since this error sum is a function of a and b a and b can be chosen so that the function has a minimum We actually use a different function 3 4 to minimize the least squares method its computations are easier and its results better understood but the results are similar m a b E ax b y 3 4 k 1 Intelligent Alarms in Anesthesia 13 The conditions to minimize this are e z b z 0 pS 0 3 5 Applying this results in m 2 ax TD y x 0 3 6a k 1 m E 2 ax b y 0 3 6b k 1 Taking in account that E 1 m the solution for a and b is m m m a 1 d mZxy Ex Yk 3 7a k 1 k 1 k 1 m m m m b 1 d mz x X XQ x E y 3 7b k 1 k 1 k 1 k 1 where m m d mEZx E x 3 8 k 1 k 1 CHE85 We use equations 3 7a and 3 8 to calculate the slope of a curve 3 2 2 3 Calculati y For an exponential curve a similar calculation is used to calculate the time constant T were we take the natural logarithm of the samples minus the offset of the exponential curve y exp TX Yo 3 9 ln y yg TX 3 10 Intelligent Alarms in Anesthesia 14 y TX 3 11 Our signals are assumed to have only the attributes described above In order to represent the signal by segments having only these attributes we must have a priori information about the signals to be
22. ave to run at the same time Task 1 Analyze signal analysis Intelligent Alarms in Anesthesia 35 Task 2 Features feature analysis Task 3 BS expert expert system Communication between the tasks is necessary because data has to flow from task 1 to task 2 and from task 2 to task 3 It is also necessary for one task to be able to control the other tasks Therefor a control queue for every task was defined each task monitors its control queue and the other tasks can put control characters in it Five message queues were defined DATA QUEUE NUMBER task 1 task 2 5 task 2 task 3 CONTROL task 1 1 task 2 task 3 3 A set of control characters that is recognized by all the tasks was defined for the control queue Abort task Suspend task Reset baselines New log file zou 3 1 Information block An information block was defined to hold all the information available about the system The information block contains the latest information on a breath to breath basis For every signal it contains some information like the name of the signal whether the monitor was calibrated or not etc It also contains the features that describe the signal with information like name value unit etc The information block is defined as follows INFORMATION BLOCK Intelligent Alarms in Anesthesia 36 SIGNAL1 name signal name available Y N monitor present and working calibrated Y N ONGOING monitor calibrat
23. ay the knowledge is implemented eg how the rules are defined 5 28 buildi i There are many expert system building tools on the market but none of them is capable of reaching a solution in a short time Speed was usually not a primary design issue In our application Intelligent Alarms in Anesthesia 29 a conclusion about the integrity of the breathing circuit has to be reached every breath eg every 6 seconds An expert system that could run in a real time environment was desired At the Eindhoven University of Technology an expert system building tool named SIMPLEXYS SIMPLe EXpert sYStem was designed by J A Blom of the Division of Medical Electrical Engineering A first version of SIMPLEXYS was available for our development SIMPLEXYS is an expert system building tool based on a semantic network although the nodes of the network are defined by rules The semantic network consists of a collection of nodes rules and relations relations that specify how a rule uses other rules A rule is either a primitive that represents an atomic concept therefore no other rules are needed in its evaluation or it is a composite a higher level concept some type of combination of other rules The rules that represent the conclusions to be evaluated are called goal rules or simply the goals Conclusions goal rules are evaluated by evaluating their constituent rules if any recursively until the recursion ends when finally the primit
24. bout the nature of the alarm situation This study is a start it focusses on the integrity of the anesthesia system If we let a computer monitor the anesthesia system in the Intelligent Alarms in Anesthesia vii same way as the anesthesiologist does and if we give the computer the same knowledge the anesthesiologist uses to reach a conclusion about the source of the alarm such an aid could be developed This report describes how a clinician monitors the anesthesia system and how to implement this knowledge in a computer so that an intelligent alarms system can be developed that aids the clinician to reach a conclusion about the possible problem more rapidly Problem approach For now we limit our intelligent alarms system to the integrity of the breathing system of the anesthesia system for a description of the anesthesia system and the breathing system see chapter 1 A definition of what can go wrong with the breathing system has to be made It has to be determined which signals have to be measured and where in the system they have to be measured A continuous analysis of these measurements has to be made to determine whether a malfunction exists and if so what exactly Symbolic data will be derived from the signals and these will be given to an expert system containing the domain specific knowledge which will reach a conclusion about the integrity of the breathing circuit The following is a list of tasks that needed to be don
25. ctrical lung model The flow is modelled with the current I the pressure with voltage V Equations I C dv dt 3 12 Intelligent Alarms in Anesthesia 16 I V V R 3 13 During inspiration the gas is forced into the patient with a constant flow I constant So V has a linear increase from eq 3 12 and V also has a linear increase because of eq 3 13 During expiration the capacitor representing mostly compliance of the lungs is discharged resulting in an exponential decrease status attribute feature 1 SLOPE up slope maximum 2 EXPON time constant 3 HORIZ minimum Table II Pressure signal cmH201 16 Figure 3 3 Normal pressure signal The pressure signal in the inspiratory hose is described in table II and in figure 3 3 3 R hac T As described in 3 4 the flow through the inspiratory hose during inspiration is constant Since only the flow in the expiratory limb is measured no flow is measured during inspiration Only when the expiratory valve is open flow can be measured When the expiratory valve opens at the beginning of Intelligent Alarms in Anesthesia 17 expiration there is a steep increase in the flow As described in 3 4 the capacitor discharges and hence the flow decreases exponentially Im1 s Figure 3 4 Normal flow signal The expiratory flow signal is described in table III and in figure 3 4 status attribute feature 1 EXPON time constant 2 HORIZ mi
26. d any rule be subsumed under another rule a SIMPLEXYS knowledge acquisition tool to automatically perform these tests is under Intelligent Alarms in Anesthesia 25 development The list of all rules is as follows BS CO2 ABSORBER CO2 INS up and FLW EX VOL unchanged and not FLW MIN down The rule uses inspired co because the co is mot scrubbed out by the co absorber The flow and pressure signals are normal INCOMP INS VALVE FLW EX VOL down and FLW MAX down and CO2 DO STR down and not FLW MIN down Expired flow and maximum fiow are down because in this case expiration takes place through both the expiratory hose and the inspiratory hose The co downstroke is prolonged because during the first part of the inspiration the gas that is already in the inspiratory hose passes the CO monitor This gas contains C05 because it is the previous exhaled gas There is no reverse flow in the expiratory hose INCOMP EXP VALVE FLW MIN down There is reverse flow through the expiratory hose because inspiration takes place through the inspiratory and expiratory hose OBST INSP HOSE PRS MAX up and PRS SLOPE up The stope of the pressure is up because the resistance of the inspiratory circuit has changed The pressure maximum is up because there is a pressure build up caused by an increased airway pressure OBST EXP HOSE FLW T CONST up and PRS T CONST up The time constants of flow and pressure are up because the resistance of t
27. d J H F Ritzerfeld M J Werter FINTTE WORDLENGTH EFFECTS IN DIGTTAL FILTERS review EUT Report 88 E 205 1988 ISBN 90 6144 205 2 Bollen M H J and G A P Jacobs EXTENS VE TESTING OF AN ALCORITHM FOR TRAVELLING WAVE BASED DIRECTIONAL DETECTION AND PHASE SELECTION BY USING TWONFIL AND EMTP EUT Report 88 E 206 1988 ISBN 90 6144 206 0 Schuurman W and M P H Weenink STABILITY OF A TAYLOR RELAXED CYLINDRICAL PLASMA SEPARATED FROM THE WALL BY A VACUUM LAYER EUT Report 88 E 207 1988 ISBN 90 6144 207 9 Lucassen F H R and H H van de Ven A NOTATION CONVENTION IN RIGID ROBOT MODELLING EUT Report 88 E 208 1988 ISBN 90 6144 208 7 J zwiak L MINIMAL REALIZATION OF SEQUENTIAL MACHINES The method of maximal adjacencies EUT Report 88 E 209 1988 ISBN 90 6144 209 5 Lucassen F H R and H H van de Ven OPTIMAL BODY FIXED COORDINATE SYSTEMS IN NEWTON EULER MODELLING EUT Report 88 E 210 1988 ISBN 90 6144 210 9 Boom A J J van den Ho CONTROL An exploratory study EUT Report 88 E 211 1988 ISBN 90 6144 211 7 Zhu Yu Cai ON THE ROBUST STABILITY OF MIMO LINEAR FEEDBACK SYSTEMS EUT Report 88 E 212 1988 ISBN 90 6144 212 5 Zhu Yu Cai M H Driessen A A H Damen and P Eykhoff A NEW SCHEME FOR IDENTIFICATION AND CONTROL EUT Report 88 E 213 1988 ISBN 90 6144 213 3 Bollen M H J and G A P Jacobs IMPLEMENTATION OF AN ALGORITHM FOR TRAVELLING WAVE BASED DIRECTIONAL DETECTION EUT Report 89 E 214 1989 ISBN
28. e 3 sizeof struct SIGNAL 5 Intelligent Alarms in Anesthesia 46 get time Summary int get time t char t pointer to time array Description This function gets the date and time as a string and stores it in array t The array will contain 26 characters and has the form Tue Jul 26 15 05 11 1988 0 0 Return Value There is no return value See Also Example get the time and date char current time 26 string lenght will always be 26 get time current time printf sin current time Intelligent Alarms in Anesthesia 47 init flag Summary int init flag flg int flg 16 bits to be zeroed Description The function init flag sets a 16 bits variable to zero This can be used for the 16 bits flag structures Return Value No return value See Also printf bin Example initialize flag structure struct CO2 FLAG flag defined in co2def h init flag flag Intelligent Alarms in Anesthesia 48 multi dos test Summary int multi dos test Description The function multi dos test tests whether the program is started up from DOS or MULTI DOS If the program was started from DOS the function prints an error message and aborts Return Value No return value See Also send features get features get control send control Example test for multi dos multi dos test test queue queue never done if no mul
29. e given in figures 4 1 to 4 6 The feature values during a malfunction have to be compared with the values in the normal situation right and left picture The graphs give a good impression of the stability of the features Intelligent Alarms in Anesthesia 23 in the normal situation Normal curves 18 85 32 20 45 50 58 85 7235 85 50 time Ico2 Eco 9 UPstr DOstr CO features Figure 4 1 CO features Normal curves 1 lg O46 4 65 1155 18 6026 20319098 6645 2551 9558 8065 2571 8078 5585200190 time max siope Tconst pressure features Figure 4 3 Pressure features Normal curves 19 45 52 80 46 19 0 46 72 75 86 10 time min max rexvoi TE Tconst low teatures Figure 4 5 Flow features Intelligent Alarms in Anesthesia Incompetent Insp valve o 10 00 2335 3785 6110 64 60 7785 0116 104 50 117 80 137 10 time ico2 Eco 9 UPstr DOsir COZ laatures Figure 4 2 CO features Incomp Insp valve oos o ee ae T N 0 9 45 23 00 3835 4070 83 10 76 40 89 65 109 00 116 35 136 05 time max 9 slope Tconst pressure lesturss ICD map valve ER E Ej E N L AL AAA 15 A a L ahhh le L helh arca ca a 1 RAN 1725 36 56 4360 7 20 70 55 83 85 97 20 110 60 130 60143 80 167 15 time mn te max 6 exl R Tconst flow taatures Figure 4 6 Flow features 24
30. e is a disconnect of the ventilator hose flow Signal flat pressure Signal flat CO signal flat FS H signal flat means that the signals have so Little variation that no breath detection could be performed 22 Intelligent Alarms in Anesthesia Leaks of three different sizes were measured Leaks of sizes 1 5mm 2mm and 3mm diameter were introduced The format of the following result table is signal FEATURE normal value 1 5mm 2mm 3mm leaks LEAK Y PIECE flow FLW EX VOL 28 gt 21 18 15 FLW MAX 38 gt 32 28 23 pressure PRS_MAX 18 16 15 14 PRS SLOPE 6 gt 4 5 4 3 PRS T CONST 0 3 gt 1 3 6 CO CO2 DO STR 100 30 40 50 CO2 DO TIME 1 5 gt 3 5 4 5 4 5 LEAK INSP HOSE flow FLW EX VOL 28 21 19 15 FLW MAX 37 32 29 25 pressure PRS T CONST 0 3 gt 2 4 6 PRS MIN 7 gt 7 7 6 PRS_MAX 18 gt 16 15 13 C05 CO2 DO STR 100 70 70 50 LEAK EXP HOSE flow FLW EX VOL 28 21 19 15 FLW MAX 37 gt 32 28 25 pressure PRS MAX 18 16 15 14 PRS T CONST 1 2 gt 2 4 6 PRS_SLOPE 6 gt 6 6 3 5 CO CO2 DO STR 100 50 50 60 CO2 DO TIME 1 8 1 8 1 8 4 2 LEAK VENT HOSE flow FLW MAX 38 gt 33 32 30 PRS EX VOL 28 gt 27 22 21 pressure PRS MAX 18 16 15 13 PRS SLOPE 6 gt 5 4 3 PRS T CONST 0 3 gt 2 4 CO CO2 DO TIME 1 5 1 5 1 5 3 5 CO2 DO STR 100 100 100 30 Examples of the graphs ar
31. e to implement an intelligent alarms system These tasks will be described in this report Define the malfunctions that we want to detect Make a comprehensive list of all dangerous malfunctions by interviewing anesthesiologists and studying the used anesthesia system chapter 2 Define which signals must be measured to detect these malfunctions and where in the system they must be measured Find out which monitors are normally used and which monitors are available Place the sensors so that an optimal malfunction detection can take place chapter 2 Record the signals that result after wilful introduction of Intelligent Alarms in Anesthesia viii a gt e Ak AA the earlier defined malfunctions with a simulated patient Sample the signals so they can be processed by a computer and plot them Analyze the recorded data define and extract their important features Define important features for every signal calculate the features chapter 3 Analyze the features define how these features change for every malfunction Extract detection rules for every malfunction chapter 4 Incorporate the derived knowledge in an expert system Make an expert system that reaches a conclusion about the integrity of the system using a set of detection rules chapter 5 A real time implementation of the intelligent alarms system has been made it is described in chapter 6 Intelligent Alarms in Anesthesia ix Chapt
32. ecommended that the I E ratio is more than 1 1 otherwise there is not enough time for exhalation LIC 74 1 4 High pressure part Two or more gases can be used with the Modulus II Usually Oxygen and Nitrous oxide are used For every gas it is possible to choose between wall supply or tank supply 1 5 Low pressure part Flow control valves set the flow of every gas The flow control valves for oxygen and nitrous oxide are mechanically interconnected This is done to be sure there is a minimum oxygen concentration of 25 in the gas mixture delivered to the patient OHM83 Vaporizers are used to add anesthetics to the gas mixture nitrous oxide is an anesthetic gas but by itself it does not provide a sufficient anesthesia An oxygen flush valve allows the anesthesiologist to temporarily give the patient an extra dose of oxygen 1 torr is almost equal to 1 mmHg 1 kPa is equal to 7 5 weg Intelligent Alarms in Anesthesia 4 1 5 1 Flow settings Not only the ventilator dials need to be set also the fresh gas flow has to be set Usually two gases are used O and N 0 So the total gas flow and the fraction of O Fi0 can be set The FiO should be such that the concentration O in the mixture is at least 21 The FiO should be as low as possible to avoid the toxicity of high concentrations of oxygen For patients with normal lungs 40 O is usually adequate but during short periods 100 can be necessary 1 6 Circle system T
33. ed cal time time when last calibrated FEATUREI name feature name unit unit of feature value value of feature valid Y N value valid or not accepted Y N valid accepted or not Stat UP DN UC static status dyn UP DN UC dynamic status PhysLowLimit lower physical limit PhysHighLimit upper physical limit FEATURE2 FEATURE3 SIGNAL2 FEATURE1 FEATURE2 The information block is filled by several tasks Every time more information becomes available the block is updated Every task can extract information from the information block The expert system for example can use the feature names to give feedback to the anesthesiologist Currently only the static status is used by the expert system but expansion of the part of the information that is used is expected 6 4 Implementation An implementation of the described techniques resulted in a demonstration prototype In this prototype the data is read out of a data file Message queue 4 is used by task 3 to signal task 1 that the run of the expert system is ready and that new TE z This information is currently not available from the monitors kk Currently not used reserved for future use Intelligent Alarms in Anesthesia 37 features can be sent This was necessary because the file read routine does not have any timing in it and the features would be overflowing the data message queue if no signal and wait were performed The prog
34. er 1 An introduction in Anesthesia 1 1 Introduction Anesthesia is a state of unconsciousness analgesia the blocking of pain and relaxation of the muscles This state is needed during surgery to help the surgeon perform the operation The anesthetic state is induced by an anesthesiologist usually a physician trained to administer anesthetics The most common method in the U S A to obtain anesthesia is inhalation of one or more anesthetic gases Because all the patient s muscles are relaxed he can not breathe himself The anesthetic mixture a mixture of anesthetic gas es and oxygen prepared by an anesthesia machine and supplied by a ventilator is forced into the patient via a breathing circuit and through a tube brought into the trachea endotracheal tube 1 2 2 hes hi There are several different anesthesia machines in use I will concentrate on the Ohmeda Modulus II Anesthesia System which is used in Shands hospital in Gainesville The Modulus II is most often used with a circle breathing System In a circle system the gases exhaled by the patient are reused A great advantage of this method is that the quantity of anesthetic agent that is needed decreases because the exhaled unused fraction of the agent is reused In the patient s lungs CO is replaced by oxygen so the patient exhales CO that has to be removed from the breathing circuit this is done by the CO absorber The following 5 parts can be identified in the
35. gy in Gainesville U S A Funding for the research was provided by Ohmeda in Madison WI manufacturer of anesthesia equipment Probl jefiniti During surgery a patient is under anesthesia This is an induced state in which the patient is unconscious insensitive to pain and the muscles of the patient are relaxed to the point where respiration must be supported by external means an anesthesia system is used for this support The risk of anesthesia is small It is estimated that between 2 and 10 anesthesia related incidents result in death for every 10 000 anesthetics ORK86 A large percentage of these incidents are due to human error and equipment failure COO78 Monitoring devices are used by the anesthesiologist for early detection of unwanted situations Most monitors are equipped with alarms that generate a sound when a user settable threshold has been exceeded At some critical moments many monitors can and will sound an alarm and the clinician is overwhelmed with an abundance of tones it is the clinician s task to analyze the multitude of unspecific alarms and reach a conclusion about what exactly is going on An unwanted situation can be the result of either an equipment malfunction or a patient problem The first thing the anesthesiologist does is to check the equipment If this appears to work well the anesthesiologist checks the patient It is our goal to develop a system that helps the anesthesiologist to reach a conclusion a
36. he circle system consists of two parts the inspiratory limb and the expiratory limb When the ventilator cycle starts the inspiratory valve opens and the expiratory valve closes Gas inspiratory inspiratory e m x hose d 7 gas C02 absorber o 94 gventilator expiratory value Figure 1 2 Circle breathing circuit Intelligent Alarms in Anesthesia 5 both fresh gas and gas recirculated through the CO absorber flows through the inspiratory hose into the patient During expiration the inspiratory valve closes and the expiratory valve opens allowing gas to flow through the expiratory hose back into the ventilator and scavenging system The patient s inflated lungs empty passively much like a balloon deflates after the ventilator has delivered the set tidal volume The CO absorber is a canister filled with soda lime which scrubs the gases passing through it from CO The soda lime has to be replaced regularly because it gets exhausted after extensive use 1 7 Scavenging system The scavenging system exists of an excess gas bag and some valves During expiration the ventilator bellows gets filled If the bellows overfills and hits the top of its case the excess gas goes into the excess gas bag If even more gas arrives a valve opens and the gas exits the system The scavenging system is needed to prevent the operation room from getting polluted with anesthetic gas mixtures Intelligent Alarms in Anesthesia
37. he expiratory circuit has changed OBST ET TUBE OBST INSP HOSE and OBST EXP HOSE According to the feature changes it appears that there is an obstruction both in the inspiratory hose and the expiratory hose DISC Y PIECE DISC PATIENT HOSE DISC EXP HOSE DISC PATIENT HOSE Intelligent Alarms in Anesthesia 26 DISC INS HOSE DISC PATIENT HOSE DISC PATIENT HOSE FLW EX VOL down and FLW MAX down and PRS flat and CO2 flat There is no breath detection on the pressure and CO signals The expiratory flow is down When the inspiratory hose is disconnected flow will only be detected if the disconnect occurs at the absorber side of the sensor DISC FGF FLW EX VOL down and FLW MAX down and PRS MAX down and PRS T CONST down Less gas and at a lower pressure will flow into the system if the fresh gas hose is disconnected DISC VENT HOSE GENERAL FAILURE There is no breath detection at all because the driving force failed There is no gas flow in the system GENERAL FAILURE CO2 flat and FLW flat and PRS flat The leak rules are currently under development Proposed rules are given below LEAK PATTERN FLW EX VOL down and FLW MAX down and PRS MAX down and PRS T CONST up and CO2 DO STR down LEAK Y PIECE LEAK PATTERN and CO2 DO TIME up and PRS SLOPE down LEAK INS HOSE LEAK PATTERN and PRS MIN down LEAK EXP HOSE LEAK Y PIECE LEAK PATTERN and PRS SLOPE down and CO2 DO TIME up LEAK VENT HOSE LEA
38. ime end time start value slope EXPON has start time end time time constant A priori information about the signal must be available to know in which phase or segment the signal is every phase has an attribute associated with it So a signal analysis routine consists of the following parts phase detection Determines in which phase the signal is parameter calculation Calculates the parameters for each phase Intelligent Alarms in Anesthesia 11 parameter validation Determines if the parameter can be calculated 3 2 1 Phase detection Phase detection is more difficult than it seems at first With well defined noise free signals it would not be difficult but data from patients is corrupted with noise and artifacts To determine the phase we can use both the signal and the derivative Limits have to be set to determine the phase This work was done by Rob Bastings BAS87 The next quote from his report describes his method of phase detection The samples are filtered with a digital low pass filter moving average filter HvO to obtain a mean value A positive and negative amplitude can be obtained by filtering the samples above and below the mean value The low and high thresholds are defined as the sum of the mean value and 50 of the positive and negative amplitude respectively This method decreases the number of false level detections An estimation of the derivative is used to determine if a high or lo
39. inhalation anesthesia system High pressure system connections to the wall outlet or gas tanks Intelligent Alarms in Anesthesia 1 Ventilatar Scavenging System Figure 1 1 Block diagram anesthesia system composed by J S Gravenstein M D Low pressure system flow control valves with flow meters vaporizer to administer anesthetics Scavenging system excess gas outlet Ventilator system ventilator hand bag Circle breathing system CO absorber hoses to connect to the patient endotracheal tube unidirectional valves The high pressure and low pressure systems together are called the anesthesia machine Intelligent Alarms in Anesthesia 2 1 3 Ventilator The ventilator is the driving source of the anesthesia system This device forces the gas mixture via the breathing circuit into the patient In case of an emergency a hand bag in the system can be used to manually ventilate the patient 1 3 1 Ventil The ventilator can be set to accommodate patients of different age weight physical condition and the type of operation Typical controls on an anesthesia ventilator are Minute volume dial This dial is used to set the number of liters per minute of gas delivered to the lung Rate dial This dial is used to set the respiratory rate number of breaths per minute I E ratio dial This dial is used to set the ratio of inspiration time to expiration time Some basic principles to set these
40. int address address of message int len number bytes in message int queue queue number 1 31 Description The function send features sends len bytes to the specified message queue starting from address Complex structures can also be send as iong as the data of these structures are continuous in the memory This function is designed for the small model of the compiler That means that all data is in one segment and that the addresses only consist of the offset An address is a pointer to a variable Return Value There is no return value See Also get features get control send control muiti dos test Example send features to queue 5 struct SIGNAL signal frame 3 defined in frame h send features signal frame 3 sizeof struct SIGNAL 5 intelligent Alarms in Anesthesia 52 suspend Summary int suspend nr int nr task number Description The function suspend stops the program and waits for the character G in its queue In this case the task number is the queve number Return Value There is no return value See Also get control send control Example suspend the program control get control 5 if control S suspend 1 Intelligent Alarms in Anesthesia 53 test queue Summary int test queue queue int queue message queue number Description The function test queue tests if there is a message in the specified que
41. ion The interrupt service routine written to get the flow pulses from the Ohmeda 5410 Volume Monitor replaces the current interrupt 0C routine COM1 is used to connect the flowmeter The interrupt routine tests status of the pin carrier detect CD by reading the modem status register If CD is low the routine increases the counter else it decreases the counter If pin 3 of connector J2 is connected to data received RD and pin 4 is connected to carrier detect CD the counter represents the volume passed since the last counter reset in amounts of 3 ml When time information is known the flow can be calculated Intelligent Alarms in Anesthesia 56 The pin connections from the Volume Monitor to a DB9 serial plug are 5410 PC DB9 1 ground J gt 5 ground 3 PULSE gt 2 RD 4 EXP FLOW gt 1 CD Literature Jourdain R L Programmer s problem solver for the IBM PC XT amp AT Englewood Cliffs N J Brady Communications Prentice Hall 1986 Ohmeda 5410 and 5420 volume monitor Service manual Madison Wis BOC Group Inc Product Service Department Ohmeda 1987 Intelligent Alarms in Anesthesia 57 Eindhoven University of Technolo Research Reports ISSN 0167 9708 Faculty of Electrical Engineering Coden TEUEDE 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 Butterweck H J an
42. ion determines in which phase segment a signal is feature update updates the current feature calculation Every sample has to be processed in order to determine in which phase a signal is and if the phase has changed Once the phase has been determined it is known from a priori knowledge which calculations have to be performed For every sample the feature currently calculated has to be updated The signal analysis part is a loop that is executed for each sample i e 20 times per second This loop consists of procedures for sample retrieval with timing provided by an A D converter the program waits until a new sample is available Intelligent Alarms in Anesthesia 33 phase detection and therefore breath detection and feature calculation 6 1 2 Feature analysis After the signal analysis has been performed and a complete breath was detected the features are available for the features analysis part see chapter 4 The detection of a breath is not straight forward Three signals are now analyzed because the sensors are not in the same place and because there is some delay between the physical input to the sensor and the electrical signal output from the sensor especially the CO monitor the completion of a breath will be detected at a different time for each signal We choose to solve this problem by waiting until in all signals a full breath is detected a breath detected flag will then be set If one signal did not detect
43. ive rules are reached This type of evaluation is called backward chaining BLO88 5 2 1 The SIMPLEXYS rule compiler The SIMPLEXYS expert system language is written in Pascal a C version is forthcoming and the rules are compiled to Pascal code by the SIMPLEXYS rule compiler Pascal procedures and variables can be defined for example to perform some action when a rule becomes true display a message control a process etc These procedures and variables are contained in the Pascal code file with the compiled rules This code file can be compiled with a Pascal compiler resulting in a fast and efficient program 2 2 2 A SIMPLEXYS program A definition of the syntax of the SIMPLEXYS expert systems language is beyond the scope of this thesis but a few concepts will be shown in the following small example program Intelligent Alarms in Anesthesia 30 line line line line DECLS type up do uc nv UP DN UC NV var flw min uc do uc nv procedure message mess text string begin writeln mess text end INITG put all the global initialization code here executed only once INITR put all the run initialization code here executed every run RULES exit exit the expert system program BTEST keypressed running The breathing circuit expert system is running STATE INITIALLY TR THEN GOAL INCOMP_EXP_VALVE INCOMP EXP VALVE There is an incompetent expiratory valve FLW_MIN_DOWN
44. like to have a sensor in every limb of our system but this would require a very complicated build up of the system with a higher possibility of errors and a higher possibility of sensor malfunction Traditionally in the Modulus II with circle breathing system the CO is measured at the Y piece pressure in the inspiratory hose and flow in the expiratory hose This is the setup we choose to use It gives us the advantage of using standard equipment and thus easier testing because we can setup our system with a standard anesthesia system Intelligent Alarms in Anesthesia 10 Chapter 3 Data analysis 3 1 Introduction If a computer program is to derive conclusions from analog signal wave forms samples of the wave forms are needed as input to the computer On practical grounds we conclude that for our signals a sample rate of 20 Hz is enough to extract the important information from these samples 3 2 General approach First we have to determine the kind of information the signals contain A set of features of each signal has to be defined so that the feature set resembles the signal s clinically useful information closely We limit our signals to almost periodic signals that can be described by a sequence of segments with the following attributes horizontal HORIZ slope SLOPE exponential curve EXPON Each attribute has parameters associated with it HORIZ has start time end time level SLOPE has start t
45. nimum 3 ee maximum Table III Flow signal Intelligent Alarms in Anesthesia 18 Chapter 4 Feature analysis 4 1 Introduction The set of features that describe all the relevant information in the signals has been defined earlier The features to be extracted are described in table IV BAS87 flow FLW MIN minimum flow Liters min FLW MAX maximum flow Liters min FLW B TIME breath time by flow sec FLW INS T inspiration time by flow sec FLW EXP T expiration time by flow sec FLW EX VOL expired volume Liters FLW T CONST time constant downstroke fldwec pressure PRS MIN minimum pressure cmH 0 PRS_MAX maximum pressure PIP L cmH 0 PRS_B TIME breath time by pressure sec PRS INS T inspiration time by pressurd sec PRS EXP T expiration time by pressure sec PRS SLOPE up slope pressure cmH50 sec PRS T CONST time constant downstr press sec CO CO2 INS inspired CO pressure level mmHg CO2 EXP expired CO pressure level mmHg CO2 B TIME breath time by CO sec CO2 DO TIME inspiration time by CO sec CO2 EXP T expiration time by CO sec CO2 UP STR CO up stroke mmHg sec CO2 DO STR CO down stroke mmHg sec Table IV Extracted features From the values of the features in this set conclusions have to be reached about the integrity of the breathing system An expert system see chapter 5 will be used in the diagnosis of the breathing system Such a system uses symbolic input of
46. on The dome is transparent and removable so a stuck valve can be confirmed and repaired Intelligent Alarms in Anesthesia 7 Dise Retainer From Absorber To Patient Figure 2 1 Inspiratory valve From the information we have about the system and from interviews with anesthesiologists we have set as our goal to automatically detect the following malfunctions Exhausted CO absorber Incompetent Inspiratory valve Incompetent Expiratory valve Obstruction Endotracheal tube Obstruction Inspiratory hose Obstruction Expiratory hose Leak Inspiratory hose Disconnect Inspiratory hose Leak Expiratory hose Disconnect Expiratory hose Leak Y piece Disconnect Y piece Leak Ventilator hose Disconnect Ventilator hose Leak CO canister Leak Endotracheal tube cuff Disconnect Endotracheal tube cuff Leak Fresh Gas Flow Disconnect Fresh Gas Flow Intelligent Alarms in Anesthesia 8 we w e In order to detect malfunctions several signals will have to be measured from the circle system If possible we would like to use the standard set of measurements 2 3 1 Monitors We want to detect malfunctions in a system where gases flow and where CO is added and extracted The logical thing to do is to measure CO pressure and flow In anesthesia monitoring CO and pressure are usually monitored but flow monitoring is not common It is much more common to measure expired volume the integral of the expiratory flow over the expiratory period
47. ool is used to implement the knowledge into a computer program The path to follow when building an expert system is straight forward Problems arise when interviewing the domain expert to extract his knowledge The expert often cannot define his knowledge in a precise unambiguous way Another problem arises when the knowledge has to be represented in a computer program There are hardly any general purpose expert system building tools available for general purpose and usually a new tool is designed for every application An expert system consists of two major parts the knowledge base and the inference engine WAT86 5 1 1 Knowledge base Most expert systems are rule based That means that the knowledge is contained in rules like if an animal has a long neck and eats leafs then it is a giraffe Advantage of this method is that rules are easy to read and easily understood Another method of implementing knowledge is a semantic net States and relations between the states are described In a semantic net it is well described how some part of the knowledge influences another part of the knowledge With semantic nets searching is optimized and checks on correctness are easier to perform 2 1 2 Inference engine The inference engine manipulates the knowledge so that a solution of the problem can be reached Some expert system building tools have a complete inference engine built in with other tools the inference process is defined by the w
48. piratory hose The rule set is currently being expanded with rules for leaks and disconnects Intelligent Alarms in Anesthesia i SAMENVATTING Abnormale en potentieel gevaarlijke foutencondities in het beademings circuit dat tijdens anesthesie gebruikt wordt bestaan onder meer uit lekken verstoppingen losgeraakte verbindingen nietwerkende kleppen en een niet goed functionerende CO absorber Omdat de huidige alarmering van de monitors niet specifiek genoeg is en onvoldoende informatie geeft is er behoefte aan een intelligent alarmerings systeem dat kan uitzoeken wat er mis is in het beademings circuit Voor de implementatie wordt gebruik gemaakt van drie signalen die gemeten worden op drie verschillende plaatsen parti le CO druk bij het T stuk druk in de inademings gedeelte van het beademings circuit gasstroom in het uitademings gedeelte van het beademings circuit Uit deze metingen werden features gehaald en de veranderingen van de features werden geanalizeerd Uit de resultaten van deze analyze werd een regelset ontworpen welke in een expert systeem geimplementeerd is m b v de SIMPLEXYS expert systeem taal Real time analyze en feature berekening zijn geimplementeerd in een multitasking omgeving Voorlopige tests zijn uitgevoerd Een implementatie van het totale systeem bleek mogelijk te zijn op een IBM AT met de multitasking omgeving van MultiDos Plus Het intelligente alarm systeem kon de volgende fouten onderscheiden niet we
49. r we have to provide an interrupt routine that replaces the current interrupt routine Because the Microsoft C compiler doesn t provide a way to write such a routine it is written in assembler Basics of an interrupt routine push the used registers on stack do the tasks you want to do DO NOT use DOS because it is not re entrant pop the registers from stack IRET instruction Because we want to use a variable in our interrupt routine that is shared by our C program we have to make sure that the data segment is set to DGROUP the name the MSC compiler uses for its small model data segment If an interrupt is invoked this is not automatically done We also have to make sure that all names that we use for segments are the same as in the assembler program See Microsoft C manual for details If the interrupt is a hardware interrupt the interrupt service routine should end with mov al 20H out 20H al This will clear the in service register so that interrupts at a lower level as the one just completed are re enabled When the interrupt service routine is written the address of the beginning of the routine has to be placed in the interrupt vector table The interrupt vector table occupies 4 bytes segment offset for every interrupt starting with interrupt 0 The begin address of the vector table is 0000 0000 The address for the interrupt service routine 0C should thus be written at 0000 0030 hexadecimal Implementat
50. rams for the three tasks were written and the communication between the tasks was made Only the SIMPLEXYS expert system is in the foreground during normal execution The expert system program also provides some control functions to control the other tasks The keyboard is monitored by the program and the other tasks can be suspended aborted etc A simple user interface is made existing of a line with available comments at the top of the screen a line at the bottom of the screen where messages can appear like suspending aborting etc The center of the screen is used to print the messages resulting from the evaluation by the expert system For demonstration purpose it was desired that the data could be shown on a screen Since MultiDos Plus only supports the low resolution CGA graphics of an IBM AT PC a second PC was used to display the curves on a high resolution EGA screen A serial communications routine for the two machines was written The samples one for every signal are sent to the serial port when they are read out of the file The other PC runs an interrupt based receive program that interrupts when a byte is received at the serial port The interrupt service routine gets the byte from the port and puts it into a ring buffer With this method no byte will ever be lost provided there is no noise on the line If three floating point values 12 bytes using the Microsoft C compiler are received the three samples are displayed 6
51. rkende inademings klep niet werkende uitademings klep uitgeputte CO absorber verstopping van het T stuk verstopping van de inademingsbuis en verstopping van de uitademingsbuis Op het moment wordt de regelset uitgebreid met regels voor lekken en onderbrekingen Intelligent Alarms in Anesthesia 11 foe ih ulace pol pie oe nh ete ey AR A ea T K STE EA CONTENTS SUMMARY 6 on n9 SAMENVATTING s e o e s e s o os oo LIST OF FIGURES s es s o LIST OF TABLES e e n INTRODUCTION ee te ew we we Chapter 1 An introduction in Anesthesia 1 1 Introduction eiue Jer Tei R m dd 1 2 Anesthesia machine ya xw telo de De 1 3 Ventilator 1 3 1 Ventilator settings 1 4 High pressure part ee 1 5 Low pressure part e 1 5 1 Flow settings 1 6 Circle system amp 5 1 7 Scavenging system Chapter 2 Signal measurement versus detection 2 1 Introduction s fed bY fa 2 2 What do we want to detect s E 2 3 Which signal do we want to measure 2 3 1 Monitors EIFE a CA dn EN v 2 3 2 Previous work gt ts Chapter 3 Data analysis s s 3 1 Introduction s a or s 3 2 General approach 10 3 2 1 Phase detection tut roe de 3 2 2 Parameter calculation P 3 2 2 1 Calculating levels 3 2 2 2 Calculating slopes 3 2 3 Parameter validation
52. s 13 Westfield Road Natick MA 01760 1986 Ohmeda Modulus II Operation manual Madison Wis Ohmeda 1983 Intelligent Alarms in Anesthesia 42 LORK86 Orkin F K Anesthetic systems In Miller R D Anaesthesia Livingstone 1986 WAT86 Waterman D A A guide to expert systems Reading Mass Addison Wesley Teknowledge series in knowledge Intelligent Alarms in Anesthesia 2nd ed New York 1986 engineering Churchill 43 APPENDIX A Programming tools The data analysis program and the feature extraction program are written in C and compiled with the Microsoft C compiler version 4 0 at this moment a change to version 5 1 is made The expert system is written in SIMPLEXYS and compiled by the SIMPLEXYS rule compiler and the Turbo Pascal compiler version 4 0 For the C programs some utility routines to be used in both programs were written The source code is contained in util c The data analysis routines are contained in anal c several include files are needed for this see fig A 1 anal c and util c were compiled and Microsofts library utility was used to put both object files in alarms lib The data analysis analyze c and the feature extraction features c have to be linked with this library co2def h padef h PLUR rame analco2 c anaiprs c and yze analflw c der nes b flagdef h flagdef h eae features bsexpert Intelligent Alarms in Anesthesia 44 APPENDI
53. spiratory hose Obstruction in the expiratory hose It is expected that leaks and disconnects can also be detected but it will be difficult to distinguish between all the leaks and disconnects New rules for the leaks and disconnects are currently under development Apart from some unexplainable crashes during startup MultiDos Plus is good way to do multitasking under the single task operating system MS DOS 7 2 Recommendations 1 The rules for the disconnect and the leaks have to be implemented Extensive tests of the intelligent alarms system have to be performed with both simulator data and real patient data Although it is not possible to introduce malfunctions while ventilating a real patient it is important to test the system in a normal situation to see how it behaves in the real violent environment of the operating room Intelligent Alarms in Anesthesia 40 2 The use of the information block has to be reconsidered it causes some overhead when it is sent from one task to another and at this moment hardly any information of the block is used It is a good way to present the information but if the information expanses it may slow down the system 3 Implementation of more sensors second flow device ventilator settings can provide a more accurate detection of malfunctions 4 The way the time constants of the flow and pressure are calculated has to be reconsidered They do not provide a very stable calculation of
54. ten with names like normal and abnormal where normal is not a numerical value but a logical one it indicates that some feature is within some range operationally defined to include all normal values In this application the following three statuses were sufficient to describe the necessary features Intelligent Alarms in Anesthesia 19 UC Unchanged The feature value is within a band around a normal value UP Up The feature value is higher than the upper threshold DN Down The feature value is lower than the lower threshold The thresholds were defined as 20 higher and 20 lower than the normal value or baseline This gives a 40 normality band This relative value the band width is expressed as a percentage gives problems if normal is close to zero Therefore a band width of 2 1 to 1 was taken if the feature s value was smaller than 5 This is useful for the CO and flow signals because their minimum is zero The values used here are arbitrary values More research needs to be done about how to derive baselines and how to set the detection bands in order to get optimum detection This research is planned for the future 4 2 Normal curves versus malfunctions In order to know how the signals or rather the features behave in the case of a malfunction we measured the three signals with the set of malfunctions of 2 2 generated by the anesthesia simulator a modified anesthesia system that can in
55. ti dos Intelligent Alarms in Anesthesia 49 printf bin Summary int printf bin flg int flag 16 bits to be printed Description The function printf bin prints a 16 bit variable as 0 s and l s No return is added This can be used to print a 16 bit flag structure Return Value No return value See Also init flag Example print a flag struct CO2 FLAG flag defined in co2def h flag bc flag to 0 printf bin flag 17 printf n 1 50 Intelligent Alarms in Anesthesia send control getcontrol Summary int send control queue c int queue queue to send to char c character to send char get control queue int queue queue to receive from Description The functions send control and get control send or get one character to or from the specified queue queue These functions are built from send features and get features Get control waits for a message if no message is available Return Value Function get control returns the character from the queue Function send control has no return value See Also send features get features test queue multi dos test Example send a character to queue 1 and get it out of there char control send control 1 A control get control 1 if control A exit 1 Inteiligent Alarms in Anesthesia 51 send features Summary int send features address len queue
56. troduce a number of malfunctions G0087 The data from these measurements were analyzed and the features were extracted with the techniques described in chapter 3 The feature values were plotted in graphs in order to compare the time course of the values that result from any single malfunction with the normal values i Normal value is the value of a feature when there are no malfunctions or disturbances A feature vatue that didn t change over a certain period of time can be called a normal value since the anesthesiologist didn t find it necessary to change something Intelligent Alarms in Anesthesia 20 The absolute changes of all features during a malfunction was obtained and listed For a list of the used abbreviations and units see table IV The format of these tables is signal FEATURE normal value malfunction value BS CO2 ABSORBER there is an exhausted CO2 absorber flow normal pressure normal CO CO2 INS P 0 gt 5 CO2 DO STR 118 111 INCOMP INS VALVE there is an incompetent inspiratory valve flow FLW_EX_VOL 36 15 FLW MAX 32 gt 16 pressure normal Co CO2_INS P 0 5 NM CO2 EXP P 48 60 slow response CO2 DO STR 118 20 CO2 UP STR 100 120 slow response INCOMP EXP VALVE there is an incompetent expiratory valve flow FLW_EXP_VOL 36 20 FLW MIN 0 11 pressure normal CO CO2 INS P 0 gt 18 CO2 EXP P 50 60 very slow response CO2 DO STR 118
57. ue This can be used if the program wants to do some other task while waiting for a message test the keyboard for a key for example Return Value Test queue returns 0 if no message is avaiable and returns 1 if there is a message in the queue See Also send features get features send control get control multi dos test Example get a message from queue 1 display keystrokes while waiting while test queue 1 if kbhit printf c getch control get_control 1 Intelligent Alarms in Anesthesia 54 APPENDIX C OMHEDA 5410 VOLUME MONITOR This paper describes how an Ohmeda 5410 Volume Monitor can be interfaced with a IBM AT XT in order to get the flow using the standard serial port RS 232 Signals According to the 5410 Service Manual page 36 the 5410 provides an Expired Flow signal at pin 3 of connector J2 The direction of the signal is given on pin 4 high is reverse flow Jumpers PJ3 and PJ4 have to be jumpered for transducer signals The signal on pin 3 will give a pulse for approximately every 3 milliliters of gas flow through the transducer Counting pulses The only thing that has to be done is to count the pulses in one period of time e g 1 second The current flow can then be calculated by flow counter 3 delta Counter is the number of pulses received in period delta e g 1 second The numeric value of the flow derived by this formula will be
58. values are Minute volume Tidal volume Respiratory rate At normal respiratory rates the tidal volume should be 10 15 ml kg body weight This is the inspired tidal volume it is delivered to the patient during the period called the inspiration time There is also an expired tidal volume It is different from the inspired tidal volume because the expired gas is at a different pressure at a different temperature has a different composition and because the respiratory quotient is not exactly equal to one Inspiratory Flow Minute volume 1 E I The inspiratory flow which is constant due to the mechanics of the system must deliver the gas volume during the Intelligent Alarms in Anesthesia 3 inspiratory part of the respiratory period it is determined by the speed at which the tidal volume is transferred from the bellows to the patient In general lower inspiratory flows produce a lower peak inspiratory pressure and lead to a better distribution of gas within the lungs The flow rate Should be adjusted to allow for an inspiratory expiratory ratio no greater than 1 in order to provide for adequate expiration Inspiratory flow rates of 40 50 L min are usually used in adult patients Alveolar respiration should be adjusted to maintain the PaCO alveolar CO level at 30 35 torr by adjustments of respiratory rate tidal volume and I E ratio Regardless of the respiratory rate and the inspiratory flow rate it is r
59. w level is reached BAS87 p 24 25 3 2 2 Parameter calculation The parameters that have to be calculated are slopes of a curve segment levels of a horizontal line and time constants of a curve segment The developed signal processing routines are robust with regard to noise i e noisy data still yield a good enough result They are not artifact resistant artifacts must be detected and removed by another method 3 2 2 1 Calculating levels In the calculation of the level of what we call a horizontal line in practice this line will be noisy and it will not always be strictly horizontal either two levels can be important the maximum and the minimum In some cases the maximum is important in other cases the minimum In the ideal Intelligent Alarms in Anesthesia 12 horizontal line the maximum and the minimum are of course the same Maximum and minimum can be calculated with the following algorithm if sample max max sample if sample min min sample A proper initialization of max and min is needed eg min highest possible minimum if the search is for a minimum 3 2 2 2 Calculating slopes Since our signal is noisy the slopes must be calculated in a way such that noisy signal values yield an accurate value for the slope Assume that our data points are modelled by the line y ax b 3 1 We want to calculate a If the k data point has an error distance from the line with m

Intelligent Alarms in Anesthesia: An implementation

Contents

Download Pdf Manuals

Related Search

Related Contents