2015 International Conference on Healthcare Informatics (ICHI) (2015)
Dallas, TX, USA
Oct. 21, 2015 to Oct. 23, 2015
DOI Bookmark: http://doi.ieeecomputersociety.org/10.1109/ICHI.2015.87
With vast amount of medical records being digitized in recent years in the form of Electronic Health Records (EHRs), accurate and large scale automated prognosis of diseases has become a possibility. However, most existing works in disease prediction have focused on a single condition or a few related conditions. Such models do not account for the fact that multiple conditions may co-occur in patients and that patient symptoms may be indicative of more than one distinct disease. Further, most learning models do not account for all patient symptoms, physiological test results, nursing notes and any other patient history that may be available in EHRs. For example, medication data, clinical notes, discharge summaries, physiological test results etc. Are all indicative of different patient conditions. Modeling such heterogeneous data simultaneously can improve automated prognosis. In this work, we propose to model such heterogeneous data simultaneously for multiple chronic disease prediction. Our prior work on multiple chronic disease prediction using clinical notes, primarily early nursing progress notes have suggested that such data are predictive and useful for prognosis. In particular, we build on the concept of admixture models . Naturally available feature splits in the form of heterogeneous patient data can be considered as 'views' of the patient. For example, a patient's discharge summary, medication, and physiological data like vital signs can be considered as three distinct views of the patient. Further, we assume that each chronic condition represents a cluster/class. Most patients suffer from one or more such conditions. Thus a patient can be considered to belong to more than one cluster. However, conventional supervised models like SVMs, logistic regression assume membership to a single class. Similarly, unsupervised models like mixture models also assume that a given data sample belongs to a single cluster. Extensions to multi-label classifications exist for SVMs that learn one SVM per-class. However, such models do not account for relationship between clusters since they treat membership to each cluster independently. Multi-task classification models  consider learning membership to each class as a separate learning task where the tasks may be related. Such models have to explicitly learn the relationships between tasks. Admixture models allow to incorporate mixed membership of samples to more than one cluster/class in a single learning task. Specifically, they impose a generative model on each view of patient data by assuming that a patient's view is generated from a probability distribution parametrized by a convex combination of parameters representing each cluster. The convex combination allows us to easily model soft-cluster membership to more than one clusters. In the case of the bag-of-words view modeled using multinomial distributions, the admixture model is equivalent to the conventional topical model Latent Dirichlet Allocation . The different views share common predictive information about chronic conditions. This similarity is captured by sharing the admixing weights across views. The admixing weights determine the convex combination of the cluster specific contribution associated with each patient. Each patient, drawn from its own admixed distribution, is parametrized using the same set of class-specific weights while generating all the patient's views. This allows us to model heterogeneous data jointly. Note that the convex combination or the weights corresponding to each patient are latent or unobserved. Our learning algorithm should determine the per-cluster parameters in each view and as well as the admixing weights shared across all views. To this end we propose to estimate the Maximum-a-Posteriori estimate of the parameters i.e. the set of parameters that maximize the posterior distribution of the parameters conditioned on the observed data. In order to incorporate supervision in this setting, we use the presence or absence of a chronic condition as an indication on the support of the convex combination. That is, we assume that, the admixing weight components corresponding to the disease that a patient is diagnosed with are non-zero and all others are 0. The precise magnitudes of the weights wherever non-zero are determined as part of our inference and learning procedure. At test time, the parameters corresponding to the cluster/class distributions in each view are fixed and the admixing weights are estimated using our inference step. Finally a suitable threshold is used to threshold the admixing weight vector per patient to determine a prognosis of all chronic diseases for the patient. Note that our inference and learning procedure converges to a local minima since the cost function is not convex. To the best of our knowledge, this is one of the first works attempting to prognose all chronic conditions simultaneously using as much digitized patient history as is possible. To summarize, we propose to predict multiple chronic conditions using heterogeneous sources of patient data available in EHRs. To model presence or absence of multiple conditions a given patient may suffer from simultaneously, we propose to model each view of patient data by leveraging admixture models. Further we propose a method to incorporate supervision in such models and to learn the view-specific parameters jointly by sharing the admixing weights across views.
Data models, Diseases, Prognostics and health management, Physiology, History, Resource management
S. Joshi, O. Koyejo and J. Ghosh, "Simultaneous Prognosis of Multiple Chronic Conditions from Heterogeneous EHR Data," 2015 International Conference on Healthcare Informatics (ICHI), Dallas, TX, USA, 2015, pp. 497.