From: AAAI Technical Report SS-94-01. Compilation copyright © 1994, AAAI (www.aaai.org). All rights reserved. Predicting risk of an adverse event in complex medical data sets using fuzzy ARTMAP network. Natalya Markuzonl,* 2, Stephan A. Gaehde2JArlene S. Ash 2Gail A. Carpenter1{ MarkA. Moskowitz IDepartment of Cognitive & Neural Systems, Boston University, 111 CummingtonSt., Boston, MA02215 2Health Care Research Unit, Section of General Internal Medicine, Evans Memorial Department of Medicine, Boston University Medical Center, Boston, MA02118 Abstract emerging patterns. Here we introduce a neural network similar to fuzzy ARTMAP (Carpenter et al., 1992). Fuzzy ARTMAP is a self-organizing Fuzzy ARTMAP is a supervised learning syspattern classification modelfor supervised learntem which includes nonlinear dynamics in the ing and prediction. The performance of the netlearning process. Weintroduce a new testing workis comparedto that of the logistic regresprocedure which allows the system to estimate sion model, using a cholecystectomy and diabetes the probability of an outcome. Simulations il- data set. The task for each model is to predict lustrate the system performance in estimating the occurrence of an adverse event based on inrisk in medical procedures. The results are com- formation about the patient provided by an input pared to the performance of the logistic regres- vector. Anadverse event is a complication followsion model. It is shown that both models have ing a medical procedure. Fuzzy ARTMAP was similar explanatory power. modified to provide probabilistic predictions, as does the logistic regression model. Performance of the two models was similar on these data. This Introduction paper includes a short description of the network architecture, the data organization, and the simThe application of neural networks to large com- ulation results. plex biomedical data sets can improve predictive performance compared to regression modeling. Neural networks can model both linear and nonFuzzy ARTMAP and probalinear relationships and update by incorporating bilistic predictions *Supportedin part hy British Petroleum(BP 89A1204). tSupportedin part by VAFellowshipTrainingGrant. Fuzzy ARTMAP (Figure 1) includes a pair tSupported in part by APd)A(ONRN00014-92-J- Fuzzy ARTmodules (ART~ and ARTb) (Carpen4015), the National Science Foundation(NSFIRI-9000530),and the ()ffice of NavalResearch(ONR N00014-ter, Grossberg and Rosen, 1991) linkedab together via an inter-ART associative memoryF that is 91-J-4100). 87 called a mapfield. During supervised learning, ART,receives a stream {a (p)} of input patterns and ARTbreceives a stream {b(P)} of patterns, where b(P) is tile correct prediction given a(P). These modulesare linked by an associative learning network and an internal controller that ensures autonomoussystem operation in real time. Tile controller is designed to create the minimal numberof ART,recognition categories needed to meet accuracy criteria. ab mapfield F ....... ......... i7 ........ match tracking ;: ARTt, ................... T wzX I rN~l Vigilance parameter p‘‘ calibrates the mini, mumconfidence that ART‘‘must have in a recognition category, or hypothesis, activated by an input a (p) in order to ART,to accept that category, I uI rather than search for a better one through an automatically controlled process of hypothesis testFigure 1: Fuzzy ARTMAP Architecture ing. Lowervalues of p‘‘ enable larger categories to form. These lower p, values lead to a broader generalization and a higher degree of (:ode com- time. This feature combined with complement pression. A predictive failure at ARTbincreases coding and fuzzy logic, lead to increasing sizes of category. Pa by the minimum amount needed to trigger hypothesis testing at ART‘‘, using a mechanism A new testing procedure is introduced which called match tracking. Matchtracking sacrifices allows the networkto, instead of predicting one of the minimumamount of generalization necessary manypossible outcomes, provide probabilities for to correct the predictive error. Hypothesis test- each outcome. During testing, instead of selecting leads to the selection of a new ART‘‘ cate- ing just one winner in the ART,, module, K segory, wh{chfocuses attention on a newcluster of lected winners combinetheir predictions to yield a(p) input features that is better able to predict a probabilistic prediction score. The influence of b(v). Match tracking allows a single ARTMAPeach category participating in the prediction is system to learn a different prediction for a rare weightedby the level of the category’s activation, event than for a cloud of similar frequent events as well as by the total numberof training patin which it is embedded. terns which participated in the formation of the The general ARTMAP network learns to clas- category. Layer G" stores the numberof training sify both binary and analogue vectors. This is inputs for each F~" category. Improvedprediction accomplished by incorporating fuzzy set-theory is achieved by training the system several times using different orderings of the input set. operations (Zadeh, 1965) into the dynamics ARTmodules which is possible due to close formal similarity between tile computations of fuzzy subsethood and ARTcategory choice, resonance, and learning. A normalization procedure called complementcoding uses on cells and off cells to represent the input pattern, and preserves the individual feature amplitudes while normalizing the total on cell/off cell vector. Learningis stable because all adaptive weights can only decrease in 88 An ARTMAP voting strategy is based on the observation that fast learlfing typically leads to different adaptive weights and recognition categories for different orderings of a given training set, even whenthe predictive accuracy of all simulations is similar. The different internal category structures cause the set of test set items where errors occur to vary from one simulation to the next. The voting strategy uses an ARTMAP system that is trained several times on one input set with different orderings. Tile final prediction for a given test set item is tile one madeby the largest numberof simulations. glucose level in the evening, is binary. Data from different patients were not combinedso that tile network was trained individually for each patient. The network was tested on three different patients. The two data sets described above offer different approaches- in the first one weare predicting the behavior of a new patient based on the exTwodata sets were tested with the network. The perience with previously observed patients, while first database was a randomsample of cholecys- in the diabetes data we make predictions based tectomy cases from 1985 and 1986, selected from on a patient’s previous history only. tile Medicare files of seven states. Abstracted data composed of tile presence or absence of about 250 key clinical findings (KCF’s) collected Results using a modified Medis(Iroups protocol (Iezzoni and Moskowitz, 1988), were provided for each of 3182 patient. Sixteen types of severe adverse Both the neural network and the logistic regresevents (including mortality within 30 days of ad- sion modelprovide a probabilistic estimate of the mission), which are major complications occur- occurrence of an adverse event while the actual ring after cholecystectomy, were defined. The oc- outcome in both data sets is binary. Model percurrence of any severe adverse event was marked formance was evaluated by the C-statistic and Ras positive in the binary outcome vector. 16.4% squared. The C-statistic maybe interpreted as of cases had at least one adverse event. The di- the likelihood that the modelcorrectly assigns a mensionality of the input vector was reduced to higher problem rank when comparing a randomly 16 from tile 62 most important KCF’s by auto- selected pair of problem/non-problemcases. The also is the area under the Receiver matic feature extraction preprocessing. KCF’s (;-statistic significantly associated (p < .05) with each of the Operating Characteristic curve which shows the 16 types of adverse events were summed.These true-positive rate against the false-positive rate 16 sums and all age variable were used to con- for a given test. R-squared is a measure of the struct all input vector. The models task was to fraction of variation in outcomethat is accounted predict tile probability of occurrence of a positive for by a given model. It is an estimate of the explanatory power of a model. outcomeafter presenting tile input vector. Data Description For the cholecystectomy database both modTile other database contained data from diabetes patients. Wepredict an abnormally high els were trained on 4/5 of the data and tested eveI~ing glucose level based on measurementsof on the remaining 1/5 of the data. Cross valipatient’s blood glucose level and insulin doses dation was performed by randomly dividing the during preceding 24 hour period. Weemployed data into 5 equal parts and successively traindata about previous evening, morningand tile af- ing models to fit each 4/5. The results of tile ternoon blood glucose level and regular and NPH neural network and tile logistic regression model insulin doses taken at these times as input in- were very similar in terms of measures provided formation to models. Thus tile temporal infor- (Table 1). (J-index for the regression model mation about changes in blood glucose level and equal to 0.68, and r "2 = 0.065. The C-index for model is equal to 0.68, and insulin doses are combined to construct a posi- the fuzzy ARTMAP 2 r = 0.062. This is a useful level of explanatory tionally organized analogue input vector. The outcome, abnormally high versus normal blood powerfor this task. 89 regression modeling ARTMAPmodeling 2C-index 2C-index R R 0.065 0.68 0.062 0.68 Table 1: database Results with tile patient 1 patient 2 patient 3 regression modeling C-index 2R 0.88 0.30 0.76 0.07 0.03 0.65 analog multidimensional maps. IEEE Transactions on Neural Networks, 3, 698-713. cholecystectomy Iezzoni, L.I., and Moskowitz, M.A. (1988). MedisGroups:a clinical assessment. JAMA, 26O, p. 3195. Zadeh, L. (1965). Fuzzy sets. Information Control, 8, 338-353. ARTMAP modeling C-index R2 0.91 0.27 0.71 0.09 0.17 0.73 Table 2: Results for different patients with the diabetes database Tile model performance for different patients in diabetes data set are very different (Table 2). This may partially be explained by physiologic variability amongpatients. Both tile logistic and neural network models were able to successfully predict the elevated blood glucose level based on previously learned variations over the preceding month. The neural network shows superior performance for patient 3, and performance similar to tile logistic: regression modelfor patient 1 and patient 2. Weconclude that modified fuzzy ARTMAP is a good alternative for estimating of risk in medical procedures. References Carpenter, G.A., Grossberg, S., and Rosen, D.B. (1991). Fuzzy ART: Fast stable learning and categorization of analog patterns by an adaptive resonance system. Neural Networks, 4, 759-771. Carpenter, G.A., Grossberg, S., Markuzon, N., Reynolds, J.H., and Rosen, D.B. (1992). Fuzzy A RTMA P: A neural network architecture for incremental supervised learning of 90