We present a recommender system aimed at improving the healthcare experience of consumers. Our model provides actionable insights to cohorts or individuals, based on their collective and personal healthrelated data. The actionable insights are delivered through digital interventions to help prevent adverse events for the consumer. By proposing timely and personalized suggestions, we will improve consumer health outcomes and prevent complications, which would also result in cost-savings. Our recommendation system employs an ensembling technique, where at its core, we have a Bayesian network that uses administrative claims data but could be extended to use Electronic Health Records (EHR) data for learning the structure of the interwoven health graph (conditions, medications, procedures, and more). This method allows for predicting the probability of various outcomes conditioned on the consumers' evidential health data. We also couple our ensemble method with a shallow random forest model to further refine the personalized recommendations after receiving the consumer's feedback. The experimental results show that our system significantly improves the precision-recall metrics of several intervention targets compared to a random baseline.
• Applied computing → Consumer health; Health care information systems; • Information systems → Recommender systems;
Availability of clinical data in form of Electronic Health Records
(EHR) has dramatically increased over the past decade. In addition
to EHRs, there are large volumes of administrative payment data that
HealthRecSys’18, October 6, 2018, Vancouver, BC, Canada
© 2018 Copyright for the individual papers remains with the authors. Copying permitted
for private and academic purposes. This volume is published and copyrighted by its
editors.
offer insights to researchers. However, administrative data and EHRs
are scattered among numerous entities and sources, such as health
plans, laboratories, providers, hospitals, chart notes, and more. In
addition to the disparate sources, the breadth, depth, linkage, and scale
of the data lead to further complexity. For end-users (consumers) it
can make interpretation difficult due to information overload or a
lack of information, as well as term inconsistency. The increasing
need to leverage the health records led to the presence of Health
Recommender Systems (HRS) [
In this paper, we present a consumer-focused recommender
system that will give individuals suggestions based on their collective
health-related data (EHR and claims). To accomplish this, we
leverage an ensemble algorithm, where a Bayesian network (BN) is
combined with a random forest (RF). Probabilistic Graphical
Models such as BN are known to tolerate the data uncertainty (noise,
ambiguity, and missing values) in a consistent and mathematically
correct way [
The structure of the remaining paper is as follows. Section 2 introduces the features of the data. Section 3 demonstrates the methods used for learning and inference. Section 4 explains the technical details of our graphical model. We discuss the experimental results in Section 5. Then, in Section 6 we review the existing techniques
Male Female and studies. We conclude the paper with a discussion and an outline of the directions for further research in Section 7. 2
The dataset used in our recommender system is largely derived from the claim data with over 1.5 million claims and over a hundred distinct attributes over a 6 year period.
Available attributes can be grouped as numerical and categorical attributes. For example, consumer gender has two states, male and female. Figure 1 illustrates that more than 55% of our target group are women. We divided Age into twelve bins as follows: ≤1, 1-4, 5-12, 13-17, 18-25, from 26 to 85 in 10-year bins, and ≥86. It was confirmed by domain experts that patients in these age ranges generally develop similar conditions. The distribution of the number of consumers per age bin is shown in Figure 2. The claim records cover 4 types of data sources: dental, medical, hospital. As it is shown in Figure 3, 65% of the claims are related to medical claims. We term the observed data, the manifests. The manifests are computed as aggregations based on meaningful categories. For each manifest, there is an integer count ≥ 0 that signifies the number of times a person had the event in a quarter. Currently, our system only relies on 4 major categories; drug classification, diagnosis classification, providers specialty, and service category. Finally, since each patient is associated with many records (claims) we aggregated records of each patient by calendar quarters (as suggested by domain experts to be the most relevant time frame to capture related health events). However, our system can be used for different time granularities. 3
We rely on a hybrid approach leveraging Probabilistic Graphical Models (PGM), Random Forest (RF), and Collaborative Filtering (CF) technique to obtain a vector of recommendations and combine the results using an ensembler. This way, we benefit from the power of PGMs in capturing the propagation of effects, CF in considering the similar situations, and RFs targeting tailored recommendations. Our proposed framework is illustrated in Figure 4 where data from different sources is fed into analysis block where we train and use our models. The output is a list of recommendations delivered via a mobile application. Based on the feedback we gather from the users (consumers, providers, etc.) on the quality of the recommendations, we can optimize the weights of the ensembler. In this study, we focus only on the green boxes, PGM and RF and the rest will be touched in our future works.
PGMs can handle large datasets in a computationally tractable manner. We leverage a Bayesian network on a Directed Acyclic Graph (DAG). Before learning the conditional probabilities of the network, we found it useful to transform the data in a way that each patient observation corresponded to the manifests exhibited in a quarter, which is long enough to cover a sequence of symptoms, that let us focus on a less noisy sequence of healthcare events. We consider populating the data for the immediate previous quarter and the next quarter for every quarter in our dataset, and hence there are triple of the observations (or rows) for any given patient. Given the transformed data, we can translate the questions to the following prediction/inference problem: Let a patient with a total of N features (manifests) including his/her medical tests, drugs, health events, and other manifests related to a period, the previous period and the next period. If we observe x features out of these N features, can we predict the values (or the probability distributions) of the remaining N −x features based on the available historical data. Using this data as input we learn the structure and create the model. The steps involved are as follows: First, to derive a joint probability distribution table, we transformed the input matrix to a discrete form with 0, 1 states. The input data had the number of times a manifest was observed for a patient in a given quarter. If this value was non-zero, we replaced it with a 1. The matrix now represents if the manifest occurred at least once in the period.
Then, we convert the matrix to have manifests observed in a quarter along with manifests observed in the next and previous quarters. This transformation provided us with the data structure with which we could predict or infer manifests of a quarter given those from another quarter. We consider the data as a matrix A. Then, we find AT × A where AT represents the transpose of A. The resulting symmetric matrix B has the number of joint occurrences of manifests across all patients. The diagonals in this matrix represent the number of occurrences of the manifest for all of the patients. Dividing the row values by the diagonal value in the row resulted in the conditional probability of the column manifest, given the row manifest, P (M R, MC) where P (M R) MR is the manifest in the row, MC is the manifest in the column, resulting in P (MC |MR). As expected, all the diagonals reduce to 1, and the resulting matrix is no more symmetric.
To determine the significant relationships and to discard those that were not as significant, we set a threshold of 0.05 (5%) for the conditional probability. If a conditional probability was greater than 5% we retain it. This choice of threshold was arbitrary, and as an improvement, we should consult with our domain experts to verify the structure and adjust accordingly. Each relationship with a conditional probability above the threshold represents a directed edge in a graph, with the arrow going from the row manifest to the column manifest. Then, we remove cyclic relationships (the diagonal entries because Bayesian model does not allow loops) using the networkX python library. This function detects cycles on a ifrst-come basis and removes the last encountered edge once a cycle is detected. These edges form a DAG structure.
The fit function estimated the Conditional Probability Distribution (CPD) for each variable based on the given data and the parameter estimation approach we use. In our case, we use the Bayesian parameter estimation because it considers the probability distribution representing our prior knowledge (how likely are we to believe in the different choices of parameters) and the support of the data (because confidence increases with more data). Moreover, our prior distribution is not uniform and hence this is also a reason to use Bayesian parameter estimation. Since our aim is to predict the values of an unknown manifest we fit the model with the training data set. At this step, the given data in graphical form was ready for performing various types of reasoning. 3.2
As shown in Figure 4, we use multiple models to obtain the final recommendation. One approach is to apply weighted ensemble methods to obtain better predictive performance than could be obtained from each of these models independently. The final recommendations can be used for a wide range of applications. Here, we focus on a Mobile App that provides health-care benefit or educational recommendations. For example, if a high probability of hypertension is predicted, then the App would recommend that the person visits his/her Primary Care Physician (PCP). 3.3
Feedback is a valuable asset for personalizing the recommendations as well as making better recommendations to similar people. The feedback loop can directly contribute to updating the weight vector of the ensemble method, as well as hyperparameter tuning of the individual models. In fact, we need an online learning algorithm to incorporate the feedback into the re-training phase. However, not every feedback or data observation has the same weight/quality. Therefore, we need to consider a context-aware algorithm to pay more attention to the more important segments of information. Attention modeling will be done as a part of our future work. 4
In this section, we briefly review the technical aspects of our Bayesian network and how parameter learning methods are used to estimate the conditional probabilities for the given set of claims and predict the state or occurrence of a new set of claims. 4.1
Bayesian network (BN) is a probabilistic graphical model (PGM) that represents a set of random variable (nodes) say X1, ....., Xn and their conditional dependencies (edges corresponding to direct influence of one node on another), say X1 ⊥⊥ X3 |X 6, using a directed acyclic graph (DAG). By surfacing these independencies we can reduce the number of values needed to be stored in order to represent the joint probability distribution and thus makes the representation more compact.
For our purpose, we use two layers of inference: structure and parameter learning. By leveraging structure inference, we create the skeleton using conditional probabilities and domain expert input which captures the dependencies between the variables. The second layer utilizes the dependencies and historical data to estimate the conditional probability distributions of the individual variables.
In parameter learning, there are two main methods: • Maximum likelihood estimation • Bayesian estimation We use Bayesian estimation over maximum likelihood estimation (MLE) because MLE considers a uniform prior distribution and this might lead us to end up in wrong conclusions about the likelihood of a variable θXi and adjust the likelihood based on whether the sample is biased or not. Also, MLE does not update the confidence of θXi with the change in the size of the data (450,000 out of 1,000,000 vs 45 out of 100). Thus, in Bayesian estimation, we use the prior knowledge about θ with its probability distribution. This distribution will represent how likely we believe the different choices of parameters. Therefore, we can create a joint distribution, which captures the assumption over the parameters θ and the data we are to observe. Each new data point gives us more information about θ and hence the probability of the next occurrence. Hence the posterior distribution for Bayesian estimation is:
Pr(θ |x [
Pr(x [
Pr(x [
Finding the conditional probability distribution (CPD) over some variables Pr(Y |E = e) is the same as inferring from a model. Therefore, predicting values for a new data point is the same as finding the conditional probability of the unknown variables, given the observed values of other variables. The CPDs can be computed from the joint probability distribution of the variables, by marginalizing and reducing them over variables and states.
In addition, we are interested in finding the state of a set of variables given other set of variables. It is simply an inference query over the model and state having higher probability would be the prediction by the model. However, computing the joint probability distribution will give us an exponentially large table, which the probabilistic graphical model helps to avoid these tables. There are two algorithms we can use for inference: and vision benefits, respectively. In each figure, the expected result from a random guess - based on the frequency of the positive class that is 27.38% (4.66%) for dental (vision) benefit - is also depicted. • Variable elimination • Belief propagation We use variable elimination over belief propagation since the former is suitable for a very large network as it is not memory-expensive. In addition, we discard the generated intermediate factors and hence it is more flexible than BP. In variable elimination, consider the model A → B → C → D and we try to find Pr(D):
Pr(D) = Õ
Pr(a) Pr(b |a) Pr(c |b) Pr(D |c) In variable elimination, we can sum over parts of the product instead of over the complete product. Hence, Eq. (2) becomes: Pr(D) =
Pr(a) Pr(b |a) Pr(c |b) Pr(D |c) = This method helps to significantly reduce the computation required to compute the probabilities. Hence variable elimination is much more efficient for calculating probability distributions than normalizing and marginalizing the joint probability distribution.
Our prediction function uses a maximum a posteriori probability to find the states of variables corresponding to the maximum probability in the joint distribution. This is useful when we want to predict the state of variables in our model. Moreover, we introduce another operation on factors called maximization. Maximum a posteriori query is essentially a way to predict the state of variables, given the state of other variables. Thus, using the trained model, we try to predict the states of variables for new data points. To design the models, we need to create conditional probability distributions or factors, add them to the base model, create an inference object, and then do maximum a posteriori queries over it for new data points to predict variable states. 5
While our un-targeted Bayesian network based recommender system can be leveraged to address a wide range of questions, in some cases a targeted model (such as a random forest) can be more beneficial. In our proposed framework depicted in Figure 4, we have included both components. To assess the abilities of our recommender system, we focus on recommendations of vision and dental benefits and train the random forest model on the same dataset and set of features that we trained our probabilistic graphical model and discuss the results in this section.
We trained the RF model using the default parameteres except for the followings: n_estimators=30,max_depth=350, random_state=0, and min_samples_leaf =2. Before fitting the data to the model we split the data into random train and test set in 80:20 ratio (widely recommended split ratio). The test set does not contain the columns/manifests that we are interested in predicting. The shape of the data used was 2,010 cases with around 1.5K features. We used a subset of the data set because of the constraint in the computational power. Figure 5 and Figure 6 illustrate the Precision-Recall ROC for dental 70 60 n o i isc50 e r P 40 30 80 70 60 iion50 s rce40 P 30 20 10
PR-ROC for Dental benefits
Model Scores
Random Guess 20 40 60 80
100
Recall PR-ROC for Vision benefits
Model Scores Random Guess
Using the trained Bayesian network, we predicted the states of all the missing columns/features of the test set. Prediction is done by belief propagation where we find the most probable state of the unknown manifests/features given the states of the other manifests/features using CPD’s. Fig 7 shows the correlation of manifests related to a specific target (Diabetes mellitus without complication, in the next quarter) and their pairwise correlations considering two consecutive quarters. As expected, having Diabetes mellitus without complication and taking antidiabetics in the previous quarter have a strong correlation with having it in the next quarter. Interestingly, the correlation between hypertension, the disorder of lipid, and high Glucose in blood are also captured by our model.
The probabilities gathered from our Bayesian network can be represented as a network. In Fig 8 the most probable manifests in the previous quarter that can be used to predict whether the consumer will visit an optimetrist are depicted as a network where edge DM.w.o.C._NXT
DM.w.o.C.
ANTIDIABETICS
METFORMIN HCL Essential hypertension
GLUCOSE BLOOD DIAGNOSTIC PRODUCTS
Disorders of lipid Other aftercare DME SUPPLIER
DME
Med. Devices
General Medicine LaboIrNaStoUrLyI N-OGuLtApRaGtiIeNnEt 1.0 thickness, shows the prevalence of manifests related to a specific target. As shown, those who had more medical interactions (surgery, medicines, ofcfie visit, etc.) are more probable to have optometry event in their medical journey.
Health-care domain has specific characteristics and requirements
that can not be addressed by general purpose or commercial
recommender systems that are available in other domains (such as ones
used in Netflix or Amazon) [
To design a framework/system, it is necessary to know the target
users. Two main end users can be considered for healthcare
recommender systems. Wiensner and Pfeifer [
To design such a system, there are several guidelines. Valdez et al.
propose a 3-step process [
There has been a number of prior efforts in this domain. One of
the well known existing application is Promedas [
Given all the efforts have been done in this domain, our model uses a large and rich dataset and relies on a mathematically correct and proved basis and is able to address a wide range of different questions about the health journey of consumers, beneficial for consumers, providers, and payers over time. 7
Our proposed recommender system provides personalized, timely and actionable health-care insights for consumers. We make relevant suggestions by predicting the probabilities of various health events. By deploying users’ feedback from their interactions within the mobile application, we enable additional personalized suggestions. This is accomplished through an ensemble algorithm, where a Bayesian network is combined with a random forest. In the future, we can improve this framework in several ways. First, by including data from other sources, such as lab results or nurse notes, we can expand the feature set to provide a more complete view of the consumer. This view would improve the precision-recall metrics of the predictions, as well as shed a brighter light on how to increase the effectiveness of the recommendations. Second, while a calendar quarter is currently the feature extraction and prediction time unit, we can change this within the probability model to predict the timing of a health event (as an additional random variable). We could also expand the model to capture a longer period of health care history to identify missing values over time, which would lead to the discovery of long-term influences, such as chronic ailments. Finally, for both better interpretability and overall improvement of the recommender system, we are working on a context-aware attention modeling algorithm to identify, invigorate, and use the most relevant features extracted from the health data and received feedback.