Software Solutions for Biomedical Decision Making Donna L. Hudson, Fellow IEEE, Maurice E. Cohen University of California, San Francisco, Fresno, CA 93701 USA
[email protected]
I. INTRODUCTION Development of usable software for biomedicine has a long history ranging from partial success to failure [1], as witnessed by the slow implementation of electronic health records, particularly in the United States [2]. Decision support systems in general have met with similar problems. While many reasons exist for this lack of success, most can be traced to a few root causes. These include the inherent complexity in medicine, lack of causative models upon which computer models can be based, and communication problems between medical professionals and software developers [3]. On the other hand, the use of electronic medical devices are wide-spread in medicine and form the basis of some of the most useful diagnostic tools, including CT and MRI imaging, ultrasound, electrocardiograms, and electroencephalo-grams [4]. E-health is a new catch phrase that is interpreted in many ways, but includes electronic medical records, telemedicine, evidence-based medicine, consumer health informatics, and health knowledge management [5]. The work described here focuses on software for medical decision making that facilitates the combination of medical data from a variety of sources to produce a complete picture of patient health. The software is based on an intelligent agent approach and incorporates several models, including a symbolic agent, a neural network model, and an approach for modeling complexity in biomedical signals. The method is illustrated in an application for differential diagnosis in cardiology.
A. Intelligent Agent Model Intelligent agents facilitate the incorporation of multiple types of reasoning paradigms, including the human decision maker. Figure 1 shows possible components in an intelligent agent model for biomedical decision making [6]. The general components in this model are: Symbolic processor (EMERGE) Data-based processor (Hypernet) Human Decision maker (Medical Professional) The agents specific to medical decision making are: Complex signal analyzer (CATS) Image analyzer (IA) B. General Purpose Agents Analysis Based on Symbolic Knowledge EMERGE, is a knowledge-based system developed by the authors for chest pain analysis [7]. It uses a modified production rule that incorporates approximate reasoning techniques with partial substantiation and weighting of antecedents. The general evidence aggregation formula is based on substantiation of proposition P determined by assuming there is some subset C of the domain V such that either the number of elements in C satisfies Q; or that each element in C satisfies the property A. The degree S to which C satisfies P is:
Task Manager
Keywords – Intelligent agents, symbolic reasoning, neural network models, biomedical decision making
II. INTELLIGENT AGENT MODEL
Agent 0 MP Agent 1 Emerge Agent 3 Hypernet
Communicator
Abstract – Software development in biomedicine continues to face challenges due to many factors including the complexity of the field, communication difficulties between healthcare professionals and software developers, and lack of standards that would facilitate communication among multiple institutions. In this work, approaches to dealing with biomedical decision making are discussed with implications in modeling the status of each patient on an individual basis. The software is based on an intelligent agent structure that incorporates a number of reasoning paradigms and a variety of input parameters ranging from text description to numerical data. The methods are illustrated in an application for cardiac diagnosis.
Agent 4 CATS Agent 5 IA
Information Pool Patient Info. Family Hx Patient Hx Phys. Findings Genetic Info. Lab Results Imaging Signal Analysis
Figure 1: Computational Architecture
S = max {VP(C)} C εA where
n VP(C) = max[(Q Σ ci^ai ) ^ min (wi ci^ai)] i=1 i=1,...,n
(1)
(2)
where ^ indicates minimum, wi and ai are the weighting factor, and degree of substantiation of the ith antecedent, ci∈{0,1}, Q is a linguistic quantifier, n is the number of antecedents. The initial rule base for EMERGE focused on evaluation of chest pain in the emergency room and has been expanded to in additional applications [8]. Analysis Based on Numeric Information Hypernet [9] is a neural network model developed by the authors and subsequently used as a classification tool in a number of medical applications, including analysis of exercise treadmill tests. Hypernet uses an expanded potential function approach in a supervised learning algorithm resulting in a nonlinear network structure with three or more layers. Input parameters can be any type of ordered data. Output is in the form of a decision function: n n n D(x) = Σ wixi + Σ Σ wijxixj i=1 i=1 j=1 (i≠j)
(3)
t-2 CTM = [ Σ δ (di)]/(t-2) i=1 where δ(di) =
(5)
1 if [(Ti+2-Ti+1)2+(Ti+1-Ti) 2].5 < r 0 otherwise.
Image Analysis Medical imaging derived using all modalities are accessible using picture archiving and computer storage (PACS) systems [11]. While many pattern recognition algorithms have been developed for automated interpretation of images, these are in general supplemental to radiologist interpretation. For this implementation, natural language descriptions of abnormalities are used. In this situation, the radiologist can also be considered an agent. III. AGENT INTERACTIONS
where xi represents the value of ith input node, wi is the weighting factor for the node, and wij is the weighting factor for the interaction of nodes i and j. If D(x) exceeds a prespecified threshold value the condition holds. The normalized value of D(x) provides a degree of certainty in the decision. Human Decision Making The human component is important in many applications and is considered here as one of the agents who can interact and react to the other agents. C. Special Purpose Agents Complex Signal Analysis System (CATS) Signal analysis data are important in many medical domains. The most widely used is the electrocardiogram in cardiology. The signal analysis agent uses a methodology based on nonlinear dynamics developed by the authors [10]. The package, denoted CATS for complex analysis of time series, utilizes Poincaré Differential Plots. For the nth point in a time series, denoted Tn the second-order difference An, is defined as: An = Tn - Tn-1
degree of variability in the time series. In order to be useful in data-based decision tools, the plot needs to be summarized numerically. This is accomplished through the use of the Central Tendency Measure (CTM) computed by selecting a circular region around the origin of radius r, counting the number of points that fall within the radius, and dividing by the total number of points. Let t = total number of points, and r = radius of central area: It is applied Holter recordings (24-hr ECG recordings).
(4)
The Poincaré Differential Plot is generated by plotting An+1 vs. An. This plot gives a useful graphical display of the
In a multi-agent system, interactions among agents become complex. Although many patterns of interaction are possible, two different paradigms are discussed here. In the first, agents act independently to arrive at a decision. At the end of the process, the decision reached by each agent, along with a reliability measure, is combined with the results from the other agents. In the second, agents share results with other agents before a decision is reached. Both approaches can be combined in the same system [12]. A. Agents Acting Independently In combining outcomes from various sources, three possibilities exist: Combine all components Choose only one component Exclude some, combine others These choices are ideally made by the automated system based on pre-defined measures, but could also be made by the human decision maker after he or she has reviewed the evidence. In choosing some components over others, some form of evidence must be available to either the automated system or the user or both. In the combination process, two types of measures are used: Reliability measure Ri i=1,..n, n=number of agents The reliability measure is an indication of the effectiveness of the method as applied to the current application. This measure takes into account the soundness of the algorithm
along with the knowledge base (symbolic or numeric) for the current application. The second factor that must be considered is not derived from the algorithm, but rather from data for the current case. Wi indicates the certainty of the result for the ith agent given the data presented for the case under evaluation. The process is illustrated in Figure 2. The combination choices are then: Select 1 Select more than 1 and combine: Equal value Weighting or each contribution B. Agents Acting Together Meta rules are used to direct a task to the appropriate methodological agent [13]. For example, a decision will be directed to the most relevant methodological agent or agents. Consider the Meta rule in Step 1 of Figure 3. Each of the three hypotheses must be evaluated. In order to determine if the ECG is abnormal, first the signal analysis analyzers must be invoked. Results are then passed to the neural network model to combine multiple parameters. The results are then passed to the Meta rule, to be combined with the other hypotheses.
# of relevant models
1
Step 1 shows the initial Meta rule, step 2 the signal analysis components, and step 3 the neural network classification. Step 3 then sends results to the Meta rule in step 1 for final evaluation. Step 2 shows two different components that contribute to the ECG evaluation: the number and type of abnormal beats and a summary measure. These numerical results are incorporated in vector x1. The classification algorithm shows D(x1) > D(x) indicating that the patient has an abnormal ECG pattern. The numerical value is translated into “presence of abnormal ECG” and sent to the knowledge-based model in step 1 where it combined with other information. Other relevant clinical parameters are evaluated by the symbolic agent to determine the type of action that is necessary. The dotted line and arrow in Figure 3 illustrates that antecedent 1 has been handled and the next task is the evaluation of antecedent 2. The task manager tracks tasks as a last-in, first-out stack (LIFO). If the method is used in a telemedicine application the blood pressure (BP) antecedent can be determined directly from a ring sensor that is monitoring remotely, and the sweating antecedent can be determined from video monitoring. IV. APPLICATIONS AND RESULTS The use of the intelligent agent system is illustrated for a cardiac diagnosis. A number of possible diagnoses are discussed, followed by a discussion of differential diagnosis.
Result, C = W1*R1
>1 Step 1: Knowledge-Based Meta Rule Model n: Max
n=n+1
>
IF
All agree?
≤ No
Yes Result, C = (ΣWi*Ri)/I
AND AND THEN
ECG abnormal BP < 150/100 Patient is perspiring Patient may be experiencing anxiety
Step 2: Signal Analysis
Ri acceptable No 1 Yes
Category
2
Arrhythmia measures, a1, a2,…
Summary measures, s1, s2,…
x1 = s1, s2,… a1, a2,…
Add to list, I=I+1 Result, C2 = (ΣWi*Ri)/I2
Result, C1 = (ΣWi*Ri)/I1
Result, C= greater of C1 , C2 Figure 2: Flowchart for Selection and Combination of Methods with Two-Category Decision
Step 3: Neural Network Classification of x1 Decision surface: D(x) = Σ wixi + ΣΣ wijxixj Weights determined by training If D(x1) > 0 ⇒ condition holds (ECG abnormal) Result returned to knowledge-based system Figure 3: Translation from Neural Network Output
A. Cardiac Diagnosis Determination of Reliability Factors The process of determining reliability factors is different for each agent and also depends on the degree of relevance of each agent to the current problem. The factors are illustrated for cardiac diagnosis. For the symbolic component, relevant considerations were discussed above. The expert knowledge for this application was acquired using a multi-year study of chest pain admissions with multiple highly qualified experts involved [14], resulting on high Likert scale values. The actual rules were derived from data in flow-chart form and include traditional conjunctive rules as well as approximate reasoning rules taking into account weighting of antecedents and partial presence of symptoms [15]. The only weak point is that the database has not been updated in depth within the last two years. The resulting Likert values are: Number of experts who were consulted Experience and reputation of each expert Relevance of expertise to topic of interest Date of last update to rule base Methods of data elicitation Average score
5 5 4 3 5 4.4
Scores are divided by 5 to normalize the range in [0, 1] giving: R1 = 0.88. For the database, all data were collected from the same population using the same criteria, so only the first three components are relevant. The database size is moderate. Amount of information Reliability of information Relevance of data to current population. Average Likert score R2 = 0.8
3 5 4 4
Quantifying results of the human decision maker can be problematic. The reliability factor is based on the level of expertise of the user but this requires self-assessment. However, each person can evaluate his or her own expertise in the specific area in question and usually arrive at a fairly accurate assessment. The human decision maker is an expert in cardiology so she assigned a reliability factor of R3 = 0.9. Determination of Weighting Factors (Symbolic Agent) Weighting factors represent the strength of the evidence in contributing to the result. For the rule-based portion, approximate reasoning techniques are used that permit the weighting of antecedents as well as the partial presence of symptoms. The rule is substantiated by aggregating evidence and comparing the result to a pre-established threshold number. The following rule illustrates the aggregation process:
IF Atrial fibrillation BP < 100/60 Abnormal mental status Cold, clammy skin Gray, cyanotic skin Weak peripheral pulses Urinary output < 30 cc/hr Continuing chest pain Heart rate > 120 THEN Patient may have MI
0.15 0.25 0.05 0.05 0.05 0.05 0.05 0.25 0.10 Threshold 0.6
d1 d2 d3 d4 d5 d6 d7 d8 d9
The rule is then evaluated using evidence aggregation as in equations (1) and (2). If the result S is greater than T, the threshold, the rule is confirmed. The weighting factor for the decision is computed as: W1= (S-T)/(1-T)
(6)
For an evidence aggregation of 0.8, the weighting factor W1 = (.9-.6)/.4 = 0.75. This is a comparative value equal to the fraction within the range of confirmation. Determination of Weighting Factor (Neural Network) For the neural network model, possible parameters are for congestive heart failure shown in Table I. The following were selected in during neural network training: x1 Abnormal Holter ECG x2 Dyspnea x3 Orthopnea x4 Edema x5 Functional impairment x6 PND x7 BUN The result is obtained by computing 7
D(x) =
7
Σ wi xi + Σ
i=1
7
Σ wi,j xixj
(7)
i=1 j=1 i≠j
Table I: Clinical parameters for Neural Network ID Bypass Dyspnea Resting HR NYHA
Age MI Orthopnea Edema
PND Rales
Gallup
Mitroreguritation
BP Total time
HR
BP
Electrolytes Drugs
HR Time to max ST NA Digitalis
Mg Ace inhib.
BUN Vasodilator
Reason for stopping Cr Antiarrhythmic
Other
URI
Demographics History Symptoms Physical findings Functional impairment ETT, Begin ETT, End
K Diruretic Viral Syndrome
Gender
where the weights have been determined using the training set. In the Hypernet neural network, the decision boundary is normalized to zero, so that positive values indicate that the condition holds and negative values indicate that the condition does not hold. In order to compute the weight W2, the value of D(x) is divided by D(y) where y is the vector with the largest positive distance in the training set [16]. Thus W2 = D(x)/D(y)= 4.6/5.4=0.85
#C #Controls CTM * Sen. Spec. Accuracy HF 32 20 X 69 91 74 25 27 X X 80 89 85 52 32 X X 83 88 85 *CTM(r=0.05), CTM(r=0.1), # of RR intervals, min r for CTM > 0.99
(8)
Determination of Weighting Factor (Human) When a person arrives at a decision, it is usually possible to estimate the degree of confidence in that decision based on the evidence provided. This estimate is represented in W3. B. Example of a Special Agent: CATS Biomedical signals play an important role in medical decision making. In cardiology, the electrocardiogram (ECG) is of prime importance and produces two important types of information: the occurrence of abnormal beats, or arrhythmias, and the long-term behavior of the signal that indicates heart rate variability. CATS addresses the latter of these. Determination of Reliability (CATS) This approach has been evaluated in a number of clinical studies to determine its effectiveness in the identification of congestive heart failure. It has been shown to be a powerful tool on its own [16]. It is even more powerful when combined with other clinical information put in results, average [17]. The reliability measure is based on an average accuracy of these clinical studies using only parameters obtained from the Holter analysis without other clinical parameters. The sensitivity, specificity, and accuracy measures are shown in Table II. R4 is computed as the average of these measures: R4 = 0.81. Determination of Weighting Factor (CATS) Figure 4 shows a Poincaré Differential Plot of a patient with congestive heart failure and a CTM value 0.24. A CTM high value indicates the absence of congestive heart failure. Thus W4 = 1- CTM = 0.76
Figure 4: Poincaré Differential Plot Patient with Congestive Heart Failure
TABLE II: Results from Holter Recordings Using CATS Analysis
C. Differential Diagnosis In the example here, different agents have expertise on different areas of heart disease. In the general case, the most likely condition would be selected as the proper diagnosis. However, an alternate approach is the use of a ranked list of possible conditions with the rankings based on the likelihood of that condition given the available information for the case. This procedure is routinely used in medicine and is denoted differential diagnosis [18]. This approach has two advantages: (1) further tests can be run to rule in or rule out conditions on the list; (2) it is possible that the patient may have multiple conditions. V. DISCUSSION AND CONCLUSION The intelligent agent approach for the development medical decision support has a number of advantages. The overall structure can be maintained for any application. The basis software structure is adapted to new applications by changing the domain knowledge. For the symbolic component, this usually requires the acquisition of a knowledge base with the help of one or more domain experts. Some attempts have been made to partially automate this process. The manner in which information is elicited is crucial. A pre-set structure may lead the expert to express knowledge in the format of the software rather than in the format of his or her usual decision making. For example, if information is sought in conjunctive form, it will lead the expert away from expressing relative weights for different parameters and from the ability to express varying levels of severity of symptoms and conditions. The challenge with data-based knowledge is different. Databases of sufficient size and consistency in data collection are needed to form a reliable learning base. In addition, the demographics of the population used for the data collection must closely match the demographics of the current patient case. For the special agents such as the signal analysis agent, the only adaptation required is based on the nature of the signal itself. For example, an ECG has a small number of channels of data and has a recurring pattern that coincides with each heartbeat. The electroencephalogram (EEG) on the other hand, has 22 or more channels of data and no recurring patterns. Although the signal analyzer works well in both cases, adjustments must be made to coincide with the goals of the analysis. In addition to cardiology, the individual methods described here have been applied in
melanoma diagnosis and prognosis [19, 20], lung cancer treatment [21], and dementia diagnosis [22]. The approach has been well-received with the medical experts with whom we have worked in these various domains. The work described here is focused on the automatic integration of the different models into a comprehensive intelligent agent model. REFERENCES [1] G Brooks, Computer assisted medical decision making, Br Med J, 29;287(6401):1299, 1983. [2] MJ Puffer, IA Ferguson, BC Wright, J Osborn, AL Anshus, BP Cahill, J Kamath, MJ Ryan, Partnering with clinical provides to enhance the efficiency of an EMR, J Healthcare Informatics & Management, 21(1)), 2007:2432. [3] DL Hudson, ME Cohen, The impact of reasoning with uncertainty in a medical expert system, Computer Applications in Medical Care, 11:43-48, 1987.s [4] H Silk, T Agresta, CM Weber, A new way to integrate clinically relevant technology into small group teaching, Acad Med, 81(3):239-44, 2006. [5] AX Gard, M Tonelli, The tension between E-health innovation and E-valuation, Arch Intern Med, 14,165(20):2329-30, 2005. [6] Hudson, D.L., Cohen, M.E., Use of Intelligent Agents in the Diagnosis of Cardiac Disorders, Computers in Cardiology, 29:633-636, 2002. [7] Hudson, D.L., Cohen, M.E., An approach to management of uncertainty in an expert system, International Journal of Intelligent Systems, 3(1):45-58, 1988. [8] ME Cohen, DL Hudson, Comparative Approaches to Medical Reasoning, World Scientific, 1995. [9] DL Hudson, ME Cohen, Neural Networks and Artificial Intelligence in Biomedical Engineering, IEEE Press/Wiley, 1999. [10] ME Cohen, DL Hudson, MF, Anderson, PC Deedwania, A conjecture to the solution of the continuous logistic equation, International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, 2(4):445461, 1994. [11] RL Arenson, SJ H Dwyer 3rd, HK Huang, Kundel, SB Seshadri, Computer applications and digital imaging, Radiology, 178(3):913-4, 1991.
[12] DL Hudson, ME Cohen, Weighing Evidence in Decision Support Systems, ISCA Computers and Their Applications, 2007, in press. [13] DL Hudson, ME Cohen, Remote Sensing and Prompting for Early Stage Dementia Patients, ISCA Software Engineering and Data Engineering, 15:116-121, 2006 [14] S Greenfield, M Nadler, The clinical management of chest pain in an emergency department: Quality assessment by criteria mapping, Medical Care, 15:898905, 1977. [15] DL Hudson, ME Cohen, An approach to management of uncertainty in an expert system, International Journal of Intelligent Systems, 3(1):45-58, 1988. [16] ME Cohen, DL Hudson, MF Anderson, PC Deedwania, Using chaotic measures to summarize medical signal analysis, IEEE Engineering in Medicine and Biology, 21:904, 1999. [17] ME Cohen, DL Hudson, New Chaotic Methods for Biomedical Signal Analysis, IEEE EMBS Information Technology Applications in Biomedicine, 117-122, 2000. [18] DL Hudson, ME Cohen, MF Anderson, PC Deedwania, Differential diagnosis using signal analysis data, IEEE Engineering in Medicine and Biology, 21:935, 1999 [19] ME Cohen, DL Hudson, PW Banda, MS Blois, Neural network approach to detection of metastatic melanoma from chromatographic analysis of urine, Computer Applications in Medical Care, 15:295-299, 1991 [20] DL Hudson, ME Cohen, PW Banda, MS Blois, M.S., Medical diagnosis and treatment plans derived from a hybrid expert system, in Hybrid Architectures for Intelligent Systems, A. Kandel and G. Langholz, Eds., CRC Press, pp. 329-344, 1992. [21] ME Cohen, DL Hudson, MF Anderson, Development of a decision making model for carcinoma of the lung, American Association for Medical Systems and Informatics, 8:164-169, 1989. [22] DL Hudson, ME Cohen, M Kramer, A Szeri, FL Chang, Diagnostic Implications of EEG Analysis in Patients with Dementia, IEEE EMBS Proceedings on Neural Engineering, 2:629-632, 2005.