Phenotyping electronic health records (EHR) focuses on defining meaningful patient groups (e.g., heart failure group and diabetes group) and identifying the temporal evolution of patients in those groups. Tensor factorization has been an effective tool for phenotyping. Most of the existing works assume either a static patient representation with aggregate data or only model temporal data. However, real EHR data contain both temporal (e.g., longitudinal clinical visits) and static information (e.g., patient demographics), which are difficult to model simultaneously. In this paper, we propose Temporal And Static TEnsor factorization (TASTE) that jointly models both static and temporal information to extract phenotypes. TASTE combines the PARAFAC2 model with non-negative matrix factorization to model a temporal and a static tensor. To fit the proposed model, we transform the original problem into simpler ones which are optimally solved in an alternating fashion. For each of the sub-problems, our proposed mathematical re-formulations lead to efficient sub-problem solvers. Comprehensive experiments on large EHR data from a heart failure (HF) study confirmed that TASTE is up to 14× faster than several baselines and the resulting phenotypes were confirmed to be clinically meaningful by a cardiologist. Using 60 phenotypes extracted by TASTE, a simple logistic regression can achieve the same level of area under the curve (AUC) for HF prediction compared to a deep learning model using recurrent neural networks (RNN) with 345 features.
Real-world predictive models in healthcare should be evaluated in terms of discrimination, the ability to differentiate between high and low risk events, and calibration, or the accuracy of the risk estimates. Unfortunately, calibration is often neglected and only discrimination is analyzed. Calibration is crucial for personalized medicine as they play an increasing role in the decision making process. Since random forest is a popular model for many healthcare applications, we propose CaliForest, a new calibrated random forest. Unlike existing calibration methodologies, CaliForest utilizes the out-of-bag samples to avoid the explicit construction of a calibration set. We evaluated CaliForest on two risk prediction tasks obtained from the publicly-available MIMIC-III database. Evaluation on these binary prediction tasks demonstrates that CaliForest can achieve the same discriminative power as random forest while obtaining a better-calibrated model evaluated across six different metrics. CaliForest is published on the standard Python software repository and the code is openly available on Github.
Necrotizing enterocolitis (NEC) is a life-threatening intestinal disease that primarily affects preterm infants during their first weeks after birth. Mortality rates associated with NEC are 15-30%, and surviving infants are susceptible to multiple serious, long-term complications. The disease is sporadic and, with currently available tools, unpredictable. We are creating an early warning system that uses stool microbiome features, combined with clinical and demographic information, to identify infants at high risk of developing NEC. Our approach uses a multiple instance learning, neural network-based system that could be used to generate daily or weekly NEC predictions for premature infants. The approach was selected to effectively utilize sparse and weakly annotated datasets characteristic of stool microbiome analysis. Here we describe initial validation of our system, using clinical and microbiome data from a nested case-control study of 161 preterm infants. We show receiver-operator curve areas above 0.9, with 75% of dominant predictive samples for NEC-affected infants identified at least 24 hours prior to disease onset. Our results pave the way for development of a real-time early warning system for NEC using a limited set of basic clinical and demographic details combined with stool microbiome data.
Adverse drug reactions (ADRs) are detrimental and unexpected clinical incidents caused by drug intake. The increasing availability of massive quantities of longitudinal event data such as electronic health records (EHRs) has redefined ADR discovery as a big data analytics problem, where data-hungry deep neural networks are especially suitable because of the abundance of the data. To this end, we introduce neural self-controlled case series (NSCCS), a deep learning framework for ADR discovery from EHRs. NSCCS rigorously follows a self-controlled case series design to adjust implicitly and efficiently for individual heterogeneity. In this way, NSCCS is robust to time-invariant confounding issues and thus more capable of identifying associations that reflect the underlying mechanism between various types of drugs and adverse conditions. We apply NSCCS to a large-scale, real-world EHR dataset and empirically demonstrate its superior performance with comprehensive experiments on a benchmark ADR discovery task.
Systematic review (SR) is an essential process to identify, evaluate, and summarize the findings of all relevant individual studies concerning health-related questions. However, conducting a SR is labor-intensive, as identifying relevant studies is a daunting process that entails multiple researchers screening thousands of articles for relevance. In this paper, we propose MMiDaS-AE, a Multi-modal Missing Data aware Stacked Autoencoder, for semi-automating screening for SRs. We use a multi-modal view that exploits three representations, of: 1) documents, 2) topics, and 3) citation networks. Documents that contain similar words will be nearby in the document embedding space. Models can also exploit the relationship between documents and the associated SR MeSH terms to capture article relevancy. Finally, related works will likely share the same citations, and thus closely related articles would, intuitively, be trained to be close to each other in the embedding space. However, using all three learned representations as features directly result in an unwieldy number of parameters. Thus, motivated by recent work on multi-modal auto-encoders, we adopt a multi-modal stacked autoencoder that can learn a shared representation encoding all three representations in a compressed space. However, in practice one or more of these modalities may be missing for an article (e.g., if we cannot recover citation information). Therefore, we propose to learn to impute the shared representation even when specific inputs are missing. We find this new model significantly improves performance on a dataset consisting of 15 SRs compared to existing approaches.
The abundance of modern health data provides many opportunities for the use of machine learning techniques to build better statistical models to improve clinical decision making. Predicting time-to-event distributions, also known as survival analysis, plays a key role in many clinical applications. We introduce a variational time-to-event prediction model, named Variational Survival Inference (VSI), which builds upon recent advances in distribution learning techniques and deep neural networks. VSI addresses the challenges of non-parametric distribution estimation by (i) relaxing the restrictive modeling assumptions made in classical models, and (ii) efficiently handling the censored observations, i.e., events that occur outside the observation window, all within the variational framework. To validate the effectiveness of our approach, an extensive set of experiments on both synthetic and real-world datasets is carried out, showing improved performance relative to competing solutions.
Machine learning models for medical image analysis often suffer from poor performance on important subsets of a population that are not identified during training or testing. For example, overall performance of a cancer detection model may be high, but the model may still consistently miss a rare but aggressive cancer subtype. We refer to this problem as hidden stratification, and observe that it results from incompletely describing the meaningful variation in a dataset. While hidden stratification can substantially reduce the clinical efficacy of machine learning models, its effects remain difficult to measure. In this work, we assess the utility of several possible techniques for measuring hidden stratification effects, and characterize these effects both via synthetic experiments on the CIFAR-100 benchmark dataset and on multiple real-world medical imaging datasets. Using these measurement techniques, we find evidence that hidden stratification can occur in unidentified imaging subsets with low prevalence, low label quality, subtle distinguishing features, or spurious correlates, and that it can result in relative performance differences of over 20% on clinically important subsets. Finally, we discuss the clinical implications of our findings, and suggest that evaluation of hidden stratification should be a critical component of any machine learning deployment in medical imaging.
The ability of caregivers and investigators to share patient data is fundamental to many areas of clinical practice and biomedical research. Prior to sharing, it is often necessary to remove identifiers such as names, contact details, and dates in order to protect patient privacy. Deidentification, the process of removing identifiers, is challenging, however. High-quality annotated data for developing models is scarce; many target identifiers are highly heterogenous (for example, there are uncountable variations of patient names); and in practice anything less than perfect sensitivity may be considered a failure. As a result, patient data is often withheld when sharing would be beneficial, and identifiable patient data is often divulged when a deidentified version would suffice. In recent years, advances in machine learning methods have led to rapid performance improvements in natural language processing tasks, in particular with the advent of large-scale pretrained language models. In this paper we develop and evaluate an approach for deidentification of clinical notes based on a bidirectional transformer model. We propose human interpretable evaluation measures and demonstrate state of the art performance against modern baseline models. Finally, we highlight current challenges in deidentification, including the absence of clear annotation guidelines, lack of portability of models, and paucity of training data. Code to develop our model is open source, allowing for broad reuse.