An attempted survey of the current state of Deep Computational Phenotyping, i.e. the use of deep learning techniques for extracting insights from large unstructured Electronic Health Records.

The goal of Computational Phenotyping is to discover descriptors of patient's health states, which can in turn be used for multiple applications, ranging from future risk identification to subject selection for clinical trials or treatment recommendation. Researchers have identified the many opportunities presented by the proliferation of Electronic Health Records systems for improving patient outcomes (the literature also uses the term Electronic Medical Records and the respective acronyms - EHRs/EMRs). This huge unstructured digital source of insights typically takes the form of diagnosis codes, medication orders, operations, and medical narratives, and even includes at times medical imagery and biomedical literature.

This page attempts to make a survey of this community and is intended for new students of the field.

I subdivide the challenges offered by the field into two domains: (1) the unique challenges offered by the processing of EMR data; (2) the challenges of doing inference in a medical setting.

• Learning without Internet-sized big data. Deep learning methods have prospered primarily in the natural language processing, computer vision, and speech recognition communities, all of which can rely on the availability of millions, hundreds of millions, of training examples. EMRs datasets have not grown to that scale yet, which makes deep learning models prone to overfitting. Many techniques have been proposed, from the careful use of regularization techniques (dropout, Bayesian priors), to transfer learning, to unsupervised weight initialization.

• Dealing with class imbalance. An extension of the prior challenge, but which takes a special importance in health care, where rare conditions affecting as low as 1% of our population of interest are the norm rather than the exception. Machine learning traditionally does poorly in tasks that face a class imbalance and the problem is further compounded in deep learning, which requires an even larger number of training examples to converge properly.

• Noise. EHR data is fundamentally noisy, sparse, incomplete, high-dimensional, and systematically biased. Several top concerns: ..* Author intent upon recording. In many cases a diagnosis code is recorded in the database to signal that laboratory tests are carried out to confirm or disprove the corresponding diagnosis, and is in no way an indication that the patient actually has the condition. ..* Medical records are a mixture of confounding interactions between disease progression and intervention processes. Informed by the current condition of a patient, physicians request medical tests which guide diagnosis, which in turn determines treatment and further testing and diagnosis. Therefore, I.I.D. assumptions, which most of our statistical tools rely on, do not hold in the face of such feedback loops and non-linear interactions. ..* Another source of noise comes from the original purpose of EHR systems - billing and avoiding liability - which may introduce a non-random bias in what is recorded and what is not. ..* When such data is used as the dependent variable, this translates to a weak label situation, given the problem of misdiagnosis, false positives, and undetected asymptomatic conditions in health care. ..* Furthermore, raw clinical events records are extremely complex. A large segment of EHR data takes the form of high-dimensional sets of discrete codes, numbering in the hundreds of thousands. Another segment is recorded through medical notes and patient narratives, which require the use of highly accurate and complex natural language processing techniques.

• Completeness. EHR data is missing in several ways: ..* Missing by mistake at random... the best-case scenario. ..* Missing by discontinuity of clinical care, in the sense that patients move among institutions, resulting in a fragmentation of data at the institutional database level. Affiliated researchers typically have access to only a short segment of a patient's total medical narrative. ..* Missing by the episodic nature of observation. Data is collected when the patient is staying the hospital, not while he is outside. ..* Missing as a result of a relationship between the patient's health status and recording. As an illustration, over the course of a hospital stay, tests are carried out more frequently when the patient's health status is changing more rapidly (as in deteriorating), whereas period of sparse observations may correspond to a health condition which is more stable over time. A reverse situation can occur when, say, an aggressive condition causes patients to rapidly die upon admission in the ER; the critically-ill patients hence leave a smaller data footprint in the EHR than the healthier cases, introducing a major bias in the analysis of the risk factors.

Conclusion: more often than not, data sparseness in EHR data is the outcome of a systematic bias and a source of information about the patient's condition, rather than the output of a stochastic process which can be easily abstracted away.

The availability of huge datasets is not the sole reason why deep learning has prospered in the NLP, computer vision and speech processing communities. I would also contend that these fields are held to far less restrictive standards than medical researchers, with regards to model interpretation, estimates of uncertainty, and causal analysis. All of these aspects continue to challenge the ability of deep learning models to this day and are areas of active research.

• Measuring uncertainty in models. Given a model and a prediction task, how certain is the model about the point estimate for that prediction? How stable are the model's parameters? Models in health care must come under scrutiny for robustness and prediction uncertainty. Such questions are still difficult to answer in deep learning, but active research is conducted on borrowing precepts from Bayesian inference, the most principled approach for quantifying uncertainty.

• Causal Analysis. Risk factors used in predictive modeling for health care are rigorously selected and tested. How do we measure the causal impact of the learned features in a neural network?

• Opening the black-box. Neural networks are still difficult to interpret. The research in determining why neural networks perform well and what they actually learn is still restricted to the field of computer vision. Creating and leveraging new methods of interpreation, visual analytics, and model diagnostics in a deep learning setting is an active area of research.

• Leveraging prior information. Deep learning techniques were developed with the idea of letting the data speak with minimal use of domain knowledge. However, this approach is wasteful in the medical setting, which has accumulated decades of expert knowledge and where we cannot rely on the huge datasets that prevail in other fields. We need to construct a general framework for including all sorts of prior information, from Bayesian-like priors informed by domain experts, to hierarchical relations among outputs and inputs, to the medical ontologies built over the last half-century, to transfering information from biomedical journals.

Fundamentally, the goal of Computational Phenotyping is to derive a concise clinically relevant representation of a patient's health state.

Such a descriptor would have countless applications:

• Automated diagnosis
• Predicting length of stay
• Cohort selection and recruitment for clinical trials
• Future risk detection
• Survival analysis and development of risk scores of morbidity and mortality
• Anomaly detection (eg. ECG vitals monitoring)
• Recommendation of treatment
• Precision medecine
• Disease progression modeling
• Unwanted readmission prediction
• Artifical assistants for physicians
• Parsing and mining of biomedical literature and medical notes and narratives.

Disclaimer: these are, in my personal opinion, the papers that addressed the challenges and applications in this field most directly, to the best of my current knowledge of the field. This list is intended for new arrivants in the field and is by no means exhaustive or finished. Papers were sorted alphabetically.

• Causal Phenotype Discovery via Deep Networks
• Computational Phenotype Discovery Using Unsupervised Feature Learning over Noisy, Sparse, and Irregular Clinical Data
• Directly Modeling Missing Data in Sequences with RNNs: Improved Classification of Clinical Time Series
• Doctor AI: Predicting Clinical Events via Recurrent Neural Networks
• Deep Computational Phenotyping
• DeepCare: A Deep Dynamic Memory Model for Predictive Medicine
• Deep Patient: An Unsupervised Representation to Predict the Future of Patients from the Electronic Health Records
• Learning to Diagnose with LSTM Recurrent Neural Networks
• Learning Robust Features using Deep Learning for Automatic Seizure Detection
• Learning Representations from EEG with Deep Recurrent Convolutional Neural Networks
• PD Disease State Assessment in Naturalistic Environments Using Deep Learning
• Unsupervised Learning of Disease Progression Models

• Multi-task Learning with Weak Class Labels: Leveraging iEEG to Detect Cortical Lesions in Cryptogenic Epilepsy
• Deep Survival Analysis
• Multi-layer Representation Learning for Medical Concepts
• Deep Survival: A Deep Cox Proportional Hazards Network
• Rubik: Knowledge Guided Tensor Factorization and Completion for Health Data Analytics
• Towards Heterogeneous Temporal Clinical Event Pattern Discovery: A Convolutional Approach
• Transferring Knowledge from Text to Predict Disease Onset
• Unsupervised Pattern Discovery in Electronic Health Care Data Using Probabilistic Clustering Models

I plan to breakdown the literature according to the challenges each paper is addressing and the kind of solutions it brings to the table. I aimed to outline here the categories and subcategories of interest (most likely based on the Challenges section above), and then place each paper into each relevant bucket (or buckets).

• Estimating uncertainty ..* There is a field of research concerned with fusing Bayesian inference with deep neural networks. Several methods have been pioneered: Expectation-Propagation, Variational Bayes. The Hamiltonian Monte Carlo and the No-U Turn Sampler seem to be the "gold standard" of MCMC methods for neural nets.
• Leveraging domain knowledge in training. ..* Weight initialization through transfer learning is common in the deep learning and can be interpreted as a form of prior knowledge. ..* Domain knowledge can also be understood as a regularizer. Some papers have looked into introducing a term to the objective function to minimize the parameters' divergence from domain expertise (as generated from, say, biomedical literature, hierarchical relationships between medical concepts, etc.)
• Model interpretation ..* Model compression, dark knowledge and mimicking: an area of research concerned with compressing the knowledge learnt in a complex (deep neural network) model into a small (interpretable) model. ..* Visualization tools for diagnosing and understanding model behavior: primarily motivated by the computer vision field. ...* T-Distributed Stochastic Neighbor Embedding (TSNE)

Most common public datasets used to test methods in the deep computational phenotyping community:

• MIMIC-II
• MIMIC-III: deidentified clinical care data collected at Beth Israel Deaconess Medical Center from 2001 to 2012. It contains over 58,000 hospital admission records of 38,645 adults and 7,875 neonates.
• Physionet challenge 2012: collection of multivariate clinical time series from 8000 ICU records. Each record is a multivariate time series of roughly 48 hours and contains 33 variables such as Albumin, heart-rate, glucose etc. The entire Physionet database

• Do a break-down by paper by theme, the challenge tackled and the solution contributed, perhaps in a table format.
• List of common freely available datasets with links used in the field and benchmarks to-date.
• List of useful workshop, conferences
• Do a full review of each paper
• There are by my estimation 40-50 papers to know in this field. I will enlarge, consolidate and sort this list.