Deep Patient: An Unsupervised Representation to Predict the Future of Patients from the Electronic Health Records

Learning Hierarchical Representations of Electronic Health Records for Clinical Outcome Prediction

Clinical outcome prediction based on Electronic Health Record (EHR) helps enable early interventions for high-risk patients, and is thus a central task for smart healthcare. Conventional deep sequential models fail to capture the rich temporal patterns encoded in the long and irregular clinical event sequences in EHR. We make the observation that clinical events at a long time scale exhibit strong temporal patterns, while events within a short time period tend to be disordered co-occurrence. We thus propose differentiated mechanisms to model clinical events at different time scales. Our model learns hierarchical representations of event sequences, to adaptively distinguish between short-range and long-range events, and accurately capture their core temporal dependencies. Experimental results on real clinical data show that our model greatly improves over previous state-of-the-art models, achieving AUC scores of 0.94 and 0.90 for predicting death and ICU admission, respectively. Our model also successfully identifies important events for different clinical outcome prediction tasks.

Interpretation of machine learning predictions for patient outcomes in electronic health records

Electronic health records are an increasingly important resource for understanding the interactions between patient health, environment, and clinical decisions. In this paper we report an empirical study of predictive modeling of seven patient outcomes using three state-of-the-art machine learning methods. Our primary goal is to validate the models by interpreting the importance of predictors in the final models. Central to interpretation is the use of feature importance scores, which vary depending on the underlying methodology. In order to assess feature importance, we compared univariate statistical tests, information-theoretic measures, permutation testing, and normalized coefficients from multivariate logistic regression models. In general we found poor correlation between methods in their assessment of feature importance, even when their performance is comparable and relatively good. However, permutation tests applied to random forest and gradient boosting models showed the most agreement, and the importance scores matched the clinical interpretation most frequently.

Wavelet entropy-based evaluation of intrinsic predictability of time series

Intrinsic predictability is imperative to quantify inherent information contained in a time series and assists in evaluating the performance of different forecasting methods to get the best possible prediction. Model forecasting performance is the measure of the probability of success. Nevertheless, model performance or the model does not provide understanding for improvement in prediction. Intuitively, intrinsic predictability delivers the highest level of predictability for a time series and informative in unfolding whether the system is unpredictable or the chosen model is a poor choice. We introduce a novel measure, the Wavelet Entropy Energy Measure (WEEM), based on wavelet transformation and information entropy for quantification of intrinsic predictability of time series. To investigate the efficiency and reliability of the proposed measure, model forecast performance was evaluated via a wavelet networks approach. The proposed measure uses the wavelet energy distribution of a time series at different scales and compares it with the wavelet energy distribution of white noise to quantify a time series as deterministic or random. We test the WEEM using a wide variety of time series ranging from deterministic, non-stationary, and ones contaminated with white noise with different noise-signal ratios. Furthermore, a relationship is developed between the WEEM and Nash-Sutcliffe Efficiency, one of the widely known measures of forecast performance. The reliability of WEEM is demonstrated by exploring the relationship to logistic map and real-world data.

Identifying Acute Low Back Pain Episodes in Primary Care Practice From Clinical Notes: Observational Study

BACKGROUND

Acute and chronic low back pain (LBP) are different conditions with different treatments. However, they are coded in electronic health records with the same International Classification of Diseases, 10th revision (ICD-10) code (M54.5) and can be differentiated only by retrospective chart reviews. This prevents an efficient definition of data-driven guidelines for billing and therapy recommendations, such as return-to-work options.

OBJECTIVE

The objective of this study was to evaluate the feasibility of automatically distinguishing acute LBP episodes by analyzing free-text clinical notes.

METHODS

We used a dataset of 17,409 clinical notes from different primary care practices; of these, 891 documents were manually annotated as acute LBP and 2973 were generally associated with LBP via the recorded ICD-10 code. We compared different supervised and unsupervised strategies for automated identification: keyword search, topic modeling, logistic regression with bag of n-grams and manual features, and deep learning (a convolutional neural network-based architecture [ConvNet]). We trained the supervised models using either manual annotations or ICD-10 codes as positive labels.

RESULTS

ConvNet trained using manual annotations obtained the best results with an area under the receiver operating characteristic curve of 0.98 and an F score of 0.70. ConvNet's results were also robust to reduction of the number of manually annotated documents. In the absence of manual annotations, topic models performed better than methods trained using ICD-10 codes, which were unsatisfactory for identifying LBP acuity.

CONCLUSIONS

This study uses clinical notes to delineate a potential path toward systematic learning of therapeutic strategies, billing guidelines, and management options for acute LBP at the point of care.

The application of unsupervised deep learning in predictive models using electronic health records

Background

The main goal of this study is to explore the use of features representing patient-level electronic health record (EHR) data, generated by the unsupervised deep learning algorithm autoencoder, in predictive modeling. Since autoencoder features are unsupervised, this paper focuses on their general lower-dimensional representation of EHR information in a wide variety of predictive tasks.

Methods

We compare the model with autoencoder features to traditional models: logistic model with least absolute shrinkage and selection operator (LASSO) and Random Forest algorithm. In addition, we include a predictive model using a small subset of response-specific variables (Simple Reg) and a model combining these variables with features from autoencoder (Enhanced Reg). We performed the study first on simulated data that mimics real world EHR data and then on actual EHR data from eight Advocate hospitals.

Results

On simulated data with incorrect categories and missing data, the precision for autoencoder is 24.16% when fixing recall at 0.7, which is higher than Random Forest (23.61%) and lower than LASSO (25.32%). The precision is 20.92% in Simple Reg and improves to 24.89% in Enhanced Reg. When using real EHR data to predict the 30-day readmission rate, the precision of autoencoder is 19.04%, which again is higher than Random Forest (18.48%) and lower than LASSO (19.70%). The precisions for Simple Reg and Enhanced Reg are 18.70 and 19.69% respectively. That is, Enhanced Reg can have competitive prediction performance compared to LASSO. In addition, results show that Enhanced Reg usually relies on fewer features under the setting of simulations of this paper.

Conclusions

We conclude that autoencoder can create useful features representing the entire space of EHR data and which are applicable to a wide array of predictive tasks. Together with important response-specific predictors, we can derive efficient and robust predictive models with less labor in data extraction and model training.

Machine learning methods in psychiatry: a brief introduction

Machine learning (ML) techniques have been widely used to address mental health questions. We discuss two main aspects of ML in psychiatry in this paper, that is, supervised learning and unsupervised learning. Examples are used to illustrate how ML has been implemented in recent mental health research.

Automated Verbal and Non-verbal Speech Analysis of Interviews of Individuals with Schizophrenia and Depression

Schizophrenia and depression are the two most common mental disorders associated with negative symptoms that contribute to poor functioning and quality of life for millions of patients globally. This study is part of a larger research project. The overall aim of the project is to develop an automated objective pipeline that aids clinical diagnosis and provides more insights into symptoms of mental illnesses. In our previous work, we have analyzed non-verbal cues and linguistic cues of individuals with schizophrenia. In this study, we extend our work to include participants with depression. Powered by natural language processing techniques, we extract verbal features, both dictionary-based and vector-based, from participants' interviews that were automatically transcribed. We also extracted conversational, phonatory, articulatory and prosodic features from the interviews to understand the conversational and acoustic characteristics of schizophrenia and depression. Combining these features, we applied ensemble learning with leave-one-out cross-validation to classify healthy controls, schizophrenic and depressive patients, achieving an accuracy of 69%-75% in paired classification. From those same features, we also predict the subjective Negative Symptoms Assessment 16 scores of patients with schizophrenia or depression, yielding an accuracy of 90.5% for NSA2 but lower accuracy for other NSA indices. Our analysis also revealed significant linguistic and non-verbal differences that are potentially symptomatic of schizophrenia and depression respectively.

Short Keynote Paper: Mainstreaming Personalized Healthcare-Transforming Healthcare Through New Era of Artificial Intelligence

Medicine has entered the digital era, driven by data from new modalities, especially genomics and imaging, as well as new sources such as wearables and Internet of Things. As we gain a deeper understanding of the disease biology and how diseases affect an individual, we are developing targeted therapies to personalize treatments. There is a need for technologies like Artificial Intelligence (AI) to be able to support predictions for personalized treatments. In order to mainstream AI in healthcare we will need to address issues such as explainability, liability and privacy. Developing explainable algorithms and including AI training in medical education are many of the solutions that can help alleviate these concerns.

A disease inference method based on symptom extraction and bidirectional Long Short Term Memory networks

The wide applications of automatic disease inference in many medical fields improve the efficiency of medical treatments. Many efforts have been made to predict patients' future health conditions according to their full clinical texts, clinical measurements or medical codes. Symptoms reflect the onset of diseases and can provide credible information for disease diagnosis. In this study, we propose a new disease inference method by extracting symptoms and integrating two symptom representation approaches. To reduce the uncertainty and irregularity of symptom descriptions in Electronic Medical Records (EMR), a comprehensive clinical knowledge database consisting of massive amount of data about diseases, symptoms, and their relationships, we extract symptoms with existing nature language process tool Metamap which is designed for biomedical texts. To take advantages of the complex relationship between symptoms and diseases to enhance the accuracy of disease inference, we present two symptom representation models: term frequency-inverse document frequency (TF-IDF) model for the representation of the relationship between symptoms and diseases and Word2Vec for the expression of the semantic relationship between symptoms. Based on these two symptom representations, we employ the bidirectional Long Short Term Memory networks (BiLSTMs) to model symptom sequences in EMR. Our proposed model shows a significant improvement in term of AUC (0.895) and F1 (0.572) for 50 diseases in MIMIC-III dataset. The results illustrate that the model with the combination of the two symptom representations perform better than the one with only one of them.

Personal and social patterns predict influenza vaccination decision

BACKGROUND

Seasonal influenza vaccination coverage remains suboptimal in most developed countries, despite longstanding recommendations of public health organizations. The individual's decision regarding vaccination is located at the core of non-adherence. We analyzed large-scale data to identify personal and social behavioral patterns for influenza vaccination uptake, and develop a model to predict vaccination decision of individuals in an upcoming influenza season.

METHODS

We analyzed primary data from the electronic medical records of a retrospective cohort of 250,000 individuals between the years 2007 and 2017, collected from 137 clinics. Individuals were randomly sampled from the database of Maccabi Healthcare Services. Maccabi's clients are representative of the Israeli population, reflect all demographic, ethnic, and socioeconomic groups and levels. We used several machine-learning models to predict whether a patient would get vaccinated in the future. Models' performance was evaluated based on the area under the ROC curve.

RESULTS

The vaccination decision of an individual can be explained in two dimensions, Personal and social. The personal dimension is strongly shaped by a "default" behavior, such as vaccination timing in previous seasons and general health consumption, but can also be affected by temporal factors such as respiratory illness in the prior year. In the social dimension, a patient is more likely to become vaccinated in a given season if at least one member of his family also became vaccinated in the same season. Vaccination uptake was highly assertive with age, socioeconomic score, and geographic location. An XGBoost-based predictive model achieved an ROC-AUC score of 0.91 with accuracy and recall rates of 90% on the test set. Prediction relied mainly on the patient's individual and household vaccination status in the past, age, number of encounters with the healthcare system, number of prescribed medications, and indicators of chronic illnesses.

CONCLUSIONS

Our ability to make an excellent prediction of the patient's decision sets a major step toward personalized influenza vaccination campaigns, and will help shape the next generation of targeted vaccination efforts.

Hierarchical Adaptive Multi-task Learning Framework for Patient Diagnoses and Diagnostic Category Classification

The problems a patient suffers from can be summarized in terms of a list of patient diagnoses. The diagnoses are typically organized in a hierarchy (or a lattice structure) in which many different low-level diagnoses are covered by one or more diagnostic categories. An interesting machine learning problem is related to learning of a wide range of diagnostic models (at different levels of abstraction) that can automatically assign a diagnosis or a diagnostic category to a specific patient. While one can always approach this problem by learning models for each diagnostic task independently, an interesting open question is how one can leverage the knowledge of a diagnostic hierarchy to improve the classification and outperform independent diagnostic models. In this work, we study this problem by designing a new hierarchical classification learning framework in which multiple diagnostic classification targets are explicitly related via diagnostic hierarchy relations. By conducting experiments on MIMIC-III data and ICD-9 diagnosis hierarchy, we demonstrate that our framework leads to improved classification performance on individual diagnostic tasks when compared to independently learned diagnostic models. This improvement is stronger for diagnoses with a low prior and smaller number of positive training examples.

Artificial intelligence approaches to predicting and detecting cognitive decline in older adults: A conceptual review

Preserving cognition and mental capacity is critical to aging with autonomy. Early detection of pathological cognitive decline facilitates the greatest impact of restorative or preventative treatments. Artificial Intelligence (AI) in healthcare is the use of computational algorithms that mimic human cognitive functions to analyze complex medical data. AI technologies like machine learning (ML) support the integration of biological, psychological, and social factors when approaching diagnosis, prognosis, and treatment of disease. This paper serves to acquaint clinicians and other stakeholders with the use, benefits, and limitations of AI for predicting, diagnosing, and classifying mild and major neurocognitive impairments, by providing a conceptual overview of this topic with emphasis on the features explored and AI techniques employed. We present studies that fell into six categories of features used for these purposes: (1) sociodemographics; (2) clinical and psychometric assessments; (3) neuroimaging and neurophysiology; (4) electronic health records and claims; (5) novel assessments (e.g., sensors for digital data); and (6) genomics/other omics. For each category we provide examples of AI approaches, including supervised and unsupervised ML, deep learning, and natural language processing. AI technology, still nascent in healthcare, has great potential to transform the way we diagnose and treat patients with neurocognitive disorders.

Precision Psychiatry Applications with Pharmacogenomics: Artificial Intelligence and Machine Learning Approaches

A growing body of evidence now suggests that precision psychiatry, an interdisciplinary field of psychiatry, precision medicine, and pharmacogenomics, serves as an indispensable foundation of medical practices by offering the accurate medication with the accurate dose at the accurate time to patients with psychiatric disorders. In light of the latest advancements in artificial intelligence and machine learning techniques, numerous biomarkers and genetic loci associated with psychiatric diseases and relevant treatments are being discovered in precision psychiatry research by employing neuroimaging and multi-omics. In this review, we focus on the latest developments for precision psychiatry research using artificial intelligence and machine learning approaches, such as deep learning and neural network algorithms, together with multi-omics and neuroimaging data. Firstly, we review precision psychiatry and pharmacogenomics studies that leverage various artificial intelligence and machine learning techniques to assess treatment prediction, prognosis prediction, diagnosis prediction, and the detection of potential biomarkers. In addition, we describe potential biomarkers and genetic loci that have been discovered to be associated with psychiatric diseases and relevant treatments. Moreover, we outline the limitations in regard to the previous precision psychiatry and pharmacogenomics studies. Finally, we present a discussion of directions and challenges for future research.

Unsupervised machine learning for the discovery of latent disease clusters and patient subgroups using electronic health records

Machine learning has become ubiquitous and a key technology on mining electronic health records (EHRs) for facilitating clinical research and practice. Unsupervised machine learning, as opposed to supervised learning, has shown promise in identifying novel patterns and relations from EHRs without using human created labels. In this paper, we investigate the application of unsupervised machine learning models in discovering latent disease clusters and patient subgroups based on EHRs. We utilized Latent Dirichlet Allocation (LDA), a generative probabilistic model, and proposed a novel model named Poisson Dirichlet Model (PDM), which extends the LDA approach using a Poisson distribution to model patients' disease diagnoses and to alleviate age and sex factors by considering both observed and expected observations. In the empirical experiments, we evaluated LDA and PDM on three patient cohorts, namely Osteoporosis, Delirium/Dementia, and Chronic Obstructive Pulmonary Disease (COPD)/Bronchiectasis Cohorts, with their EHR data retrieved from the Rochester Epidemiology Project (REP) medical records linkage system, for the discovery of latent disease clusters and patient subgroups. We compared the effectiveness of LDA and PDM in identifying disease clusters through the visualization of disease representations. We tested the performance of LDA and PDM in differentiating patient subgroups through survival analysis, as well as statistical analysis of demographics and Elixhauser Comorbidity Index (ECI) scores in those subgroups. The experimental results show that the proposed PDM could effectively identify distinguished disease clusters based on the latent patterns hidden in the EHR data by alleviating the impact of age and sex, and that LDA could stratify patients into differentiable subgroups with larger p-values than PDM. However, those subgroups identified by LDA are highly associated with patients' age and sex. The subgroups discovered by PDM might imply the underlying patterns of diseases of greater interest in epidemiology research due to the alleviation of age and sex. Both unsupervised machine learning approaches could be leveraged to discover patient subgroups using EHRs but with different foci.

Systems Biology and Kidney Disease

The kidney is a complex organ responsible for maintaining multiple aspects of homeostasis in the human body. The combination of distinct, yet interrelated, molecular functions across different cell types make the delineation of factors associated with loss or decline in kidney function challenging. Consequently, there has been a paucity of new diagnostic markers and treatment options becoming available to clinicians and patients in managing kidney diseases. A systems biology approach to understanding the kidney leverages recent advances in computational technology and methods to integrate diverse sets of data. It has the potential to unravel the interplay of multiple genes, proteins, and molecular mechanisms that drive key functions in kidney health and disease. The emergence of large, detailed, multilevel biologic and clinical data from national databases, cohort studies, and trials now provide the critical pieces needed for meaningful application of systems biology approaches in nephrology. The purpose of this review is to provide an overview of the current state in the evolution of the field. Recent successes of systems biology to identify targeted therapies linked to mechanistic biomarkers in the kidney are described to emphasize the relevance to clinical care and the outlook for improving outcomes for patients with kidney diseases.

Deep Learning for Natural Language Processing in Radiology-Fundamentals and a Systematic Review

PURPOSE

Natural language processing (NLP) enables conversion of free text into structured data. Recent innovations in deep learning technology provide improved NLP performance. We aimed to survey deep learning NLP fundamentals and review radiology-related research.

METHODS

This systematic review was reported according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines. We searched for deep learning NLP radiology studies published up to September 2019. MEDLINE, Scopus, and Google Scholar were used as search databases.

RESULTS

Ten relevant studies published between 2018 and 2019 were identified. Deep learning models applied for NLP in radiology are convolutional neural networks, recurrent neural networks, long short-term memory networks, and attention networks. Deep learning NLP applications in radiology include flagging of diagnoses such as pulmonary embolisms and fractures, labeling follow-up recommendations, and automatic selection of imaging protocols. Deep learning NLP models perform as well as or better than traditional NLP models.

CONCLUSION

Research and use of deep learning NLP in radiology is increasing. Acquaintance with this technology can help prepare radiologists for the coming changes in their field.

Benchmarking Deep Learning Architectures for Predicting Readmission to the ICU and Describing Patients-at-Risk

To compare different deep learning architectures for predicting the risk of readmission within 30 days of discharge from the intensive care unit (ICU). The interpretability of attention-based models is leveraged to describe patients-at-risk. Several deep learning architectures making use of attention mechanisms, recurrent layers, neural ordinary differential equations (ODEs), and medical concept embeddings with time-aware attention were trained using publicly available electronic medical record data (MIMIC-III) associated with 45,298 ICU stays for 33,150 patients. Bayesian inference was used to compute the posterior over weights of an attention-based model. Odds ratios associated with an increased risk of readmission were computed for static variables. Diagnoses, procedures, medications, and vital signs were ranked according to the associated risk of readmission. A recurrent neural network, with time dynamics of code embeddings computed by neural ODEs, achieved the highest average precision of 0.331 (AUROC: 0.739, F₁-Score: 0.372). Predictive accuracy was comparable across neural network architectures. Groups of patients at risk included those suffering from infectious complications, with chronic or progressive conditions, and for whom standard medical care was not suitable. Attention-based networks may be preferable to recurrent networks if an interpretable model is required, at only marginal cost in predictive accuracy.

Outlier Detection in Health Record Free-Text using Deep Learning

In recent years, machine learning approaches have been successfully applied to analysis of patient symptom data in the context of disease diagnosis, at least where such data is well codified. However, much of the data present in Electronic Health Records (EHR) is unlikely to prove suitable for classic machine learning approaches. In particular, the use of free (or unstructured) text for clinical notes presents significant analytical opportunities, but also unique difficulties. Furthermore, the wide dispersal of health data relating to individuals necessitates the development of decentralized solutions. We provide, in this paper, an overview of our approach to develop a neural network framework for patient classification in the environment of EHRs where data may be heterogeneous, incomplete (containing missing values), and noisy. In this paper we describe our system which provides prediction of outlier cases which are likely to relate to frequent attender patients, which acheives an Area-Under-the-Curve score of up to 0.92.

Prediction of gestational diabetes based on nationwide electronic health records

Gestational diabetes mellitus (GDM) poses increased risk of short- and long-term complications for mother and offspring^1-4. GDM is typically diagnosed at 24-28 weeks of gestation, but earlier detection is desirable as this may prevent or considerably reduce the risk of adverse pregnancy outcomes^5,6. Here we used a machine-learning approach to predict GDM on retrospective data of 588,622 pregnancies in Israel for which comprehensive electronic health records were available. Our models predict GDM with high accuracy even at pregnancy initiation (area under the receiver operating curve (auROC) = 0.85), substantially outperforming a baseline risk score (auROC = 0.68). We validated our results on both a future validation set and a geographical validation set from the most populated city in Israel, Jerusalem, thereby emulating real-world performance. Interrogating our model, we uncovered previously unreported risk factors, including results of previous pregnancy glucose challenge tests. Finally, we devised a simpler model based on just nine questions that a patient could answer, with only a modest reduction in accuracy (auROC = 0.80). Overall, our models may allow early-stage intervention in high-risk women, as well as a cost-effective screening approach that could avoid the need for glucose tolerance tests by identifying low-risk women. Future prospective studies and studies on additional populations are needed to assess the real-world clinical utility of the model.

Therapeutic target database 2020: enriched resource for facilitating research and early development of targeted therapeutics

Knowledge of therapeutic targets and early drug candidates is useful for improved drug discovery. In particular, information about target regulators and the patented therapeutic agents facilitates research regarding druggability, systems pharmacology, new trends, molecular landscapes, and the development of drug discovery tools. To complement other databases, we constructed the Therapeutic Target Database (TTD) with expanded information about (i) target-regulating microRNAs and transcription factors, (ii) target-interacting proteins, and (iii) patented agents and their targets (structures and experimental activity values if available), which can be conveniently retrieved and is further enriched with regulatory mechanisms or biochemical classes. We also updated the TTD with the recently released International Classification of Diseases ICD-11 codes and additional sets of successful, clinical trial, and literature-reported targets that emerged since the last update. TTD is accessible at http://bidd.nus.edu.sg/group/ttd/ttd.asp. In case of possible web connectivity issues, two mirror sites of TTD are also constructed (http://db.idrblab.org/ttd/ and http://db.idrblab.net/ttd/).

Deep learning in omics: a survey and guideline

Omics, such as genomics, transcriptome and proteomics, has been affected by the era of big data. A huge amount of high dimensional and complex structured data has made it no longer applicable for conventional machine learning algorithms. Fortunately, deep learning technology can contribute toward resolving these challenges. There is evidence that deep learning can handle omics data well and resolve omics problems. This survey aims to provide an entry-level guideline for researchers, to understand and use deep learning in order to solve omics problems. We first introduce several deep learning models and then discuss several research areas which have combined omics and deep learning in recent years. In addition, we summarize the general steps involved in using deep learning which have not yet been systematically discussed in the existent literature on this topic. Finally, we compare the features and performance of current mainstream open source deep learning frameworks and present the opportunities and challenges involved in deep learning. This survey will be a good starting point and guideline for omics researchers to understand deep learning.

Era of Intelligent Systems in Healthcare

The aim of this chapter is to prepare the reader for the outstanding trip that she/he embarked when starting reading this book. At first, we shall try to look for answers to some of the most important questions regarding the connection between intelligent systems and healthcare. What are intelligent systems? How can they be used in healthcare? Have they got benefits and prospects? Let us highlight some of the decisive factors for a successful deployment of intelligent systems in healthcare, including intelligent clinical support and intelligent patient management.

Monitoring ICU Mortality Risk with A Long Short-Term Memory Recurrent Neural Network

In intensive care units (ICU), mortality prediction is a critical factor not only for effective medical intervention but also for allocation of clinical resources. Structured electronic health records (EHR) contain valuable information for assessing mortality risk in ICU patients, but current mortality prediction models usually require laborious human-engineered features. Furthermore, substantial missing data in EHR is a common problem for both the construction and implementation of a prediction model.Inspired by language-related models, we design a new framework for dynamic monitoring of patients' mortality risk. Our framework uses the bag-of-words representation for all relevant medical events based on most recent history as inputs. By design, it is robust to missing data in EHR and can be easily implemented as an instant scoring system to monitor the medical development of all ICU patients. Specifically, our model uses latent semantic analysis (LSA) to encode the patients' states into low-dimensional embeddings, which are further fed to long short-term memory networks for mortality risk prediction. Our results show that the deep learning based framework performs better than the existing severity scoring system, SAPS-II. We observe that bidirectional long short-term memory demonstrates superior performance, probably due to the successful capture of both forward and backward temporal dependencies.

Representation learning for clinical time series prediction tasks in electronic health records

Background

Electronic health records (EHRs) provide possibilities to improve patient care and facilitate clinical research. However, there are many challenges faced by the applications of EHRs, such as temporality, high dimensionality, sparseness, noise, random error and systematic bias. In particular, temporal information is difficult to effectively use by traditional machine learning methods while the sequential information of EHRs is very useful.

Method

In this paper, we propose a general-purpose patient representation learning approach to summarize sequential EHRs. Specifically, a recurrent neural network based denoising autoencoder (RNN-DAE) is employed to encode inhospital records of each patient into a low dimensional dense vector.

Results

Based on EHR data collected from Shuguang Hospital affiliated to Shanghai University of Traditional Chinese Medicine, we experimentally evaluate our proposed RNN-DAE method on both mortality prediction task and comorbidity prediction task. Extensive experimental results show that our proposed RNN-DAE method outperforms existing methods. In addition, we apply the “Deep Feature” represented by our proposed RNN-DAE method to track similar patients with t-SNE, which also achieves some interesting observations.

Conclusion

We propose an effective unsupervised RNN-DAE method to summarize patient sequential information in EHR data. Our proposed RNN-DAE method is useful on both mortality prediction task and comorbidity prediction task.

Multi-objective ensemble deep learning using electronic health records to predict outcomes after lung cancer radiotherapy

Accurately predicting treatment outcome is crucial for creating personalized treatment plans and follow-up schedules. Electronic health records (EHRs) contain valuable patient-specific information that can be leveraged to improve outcome prediction. We propose a reliable multi-objective ensemble deep learning (MoEDL) method that uses features extracted from EHRs to predict high risk of treatment failure after radiotherapy in patients with lung cancer. The dataset used in this study contains EHRs of 814 patients who had not achieved disease-free status and 193 patients who were disease-free with at least one year follow-up time after lung cancer radiation therapy. The proposed MoEDL consists of three phases: (1) training with dynamic ensemble deep learning; (2) model selection with adaptive multi-objective optimization; and (3) testing with evidential reasoning (ER) fusion. Specifically, in the training phase, we employ deep perceptron networks as base learners to handle various issues with EHR data. The architecture and key hyper-parameters of each base learner are dynamically adjusted to increase the diversity of learners while reducing the time spent tuning hyper-parameters. Furthermore, we integrate the snapshot ensembles (SE) restarting strategy, multi-objective optimization, and ER fusion to improve the prediction robustness and accuracy of individual networks. The SE restarting strategy can yield multiple candidate models at no additional training cost in the training stage. The multi-objective model simultaneously considers sensitivity, specificity, and AUC as objective functions, overcoming the limitations of single-objective-based model selection. For the testing stage, we utilized an analytic ER rule to fuse the output scores from each optimal model to obtain reliable and robust predictive results. Our experimental results demonstrate that MoEDL can perform better than other conventional methods.

Prediction of health care expenditure increase: how does pharmacotherapy contribute?

BACKGROUND

Rising health care costs are a major public health issue. Thus, accurately predicting future costs and understanding which factors contribute to increases in health care expenditures are important. The objective of this project was to predict patients healthcare costs development in the subsequent year and to identify factors contributing to this prediction, with a particular focus on the role of pharmacotherapy.

METHODS

We used 2014-2015 Swiss health insurance claims data on 373'264 adult patients to classify individuals' changes in health care costs. We performed extensive feature generation and developed predictive models using logistic regression, boosted decision trees and neural networks. Based on the decision tree model, we performed a detailed feature importance analysis and subgroup analysis, with an emphasis on drug classes.

RESULTS

The boosted decision tree model achieved an overall accuracy of 67.6% and an area under the curve-score of 0.74; the neural network and logistic regression models performed 0.4 and 1.9% worse, respectively. Feature engineering played a key role in capturing temporal patterns in the data. The number of features was reduced from 747 to 36 with only a 0.5% loss in the accuracy. In addition to hospitalisation and outpatient physician visits, 6 drug classes and the mode of drug administration were among the most important features. Patient subgroups with a high probability of increase (up to 88%) and decrease (up to 92%) were identified.

CONCLUSIONS

Pharmacotherapy provides important information for predicting cost increases in the total population. Moreover, its relative importance increases in combination with other features, including health care utilisation.

Variational Autoencoders for Cancer Data Integration: Design Principles and Computational Practice

International initiatives such as the Molecular Taxonomy of Breast Cancer International Consortium are collecting multiple data sets at different genome-scales with the aim to identify novel cancer bio-markers and predict patient survival. To analyze such data, several machine learning, bioinformatics, and statistical methods have been applied, among them neural networks such as autoencoders. Although these models provide a good statistical learning framework to analyze multi-omic and/or clinical data, there is a distinct lack of work on how to integrate diverse patient data and identify the optimal design best suited to the available data.In this paper, we investigate several autoencoder architectures that integrate a variety of cancer patient data types (e.g., multi-omics and clinical data). We perform extensive analyses of these approaches and provide a clear methodological and computational framework for designing systems that enable clinicians to investigate cancer traits and translate the results into clinical applications. We demonstrate how these networks can be designed, built, and, in particular, applied to tasks of integrative analyses of heterogeneous breast cancer data. The results show that these approaches yield relevant data representations that, in turn, lead to accurate and stable diagnosis.

Linking glycemic dysregulation in diabetes to symptoms, comorbidities, and genetics through EHR data mining

Diabetes is a diverse and complex disease, with considerable variation in phenotypic manifestation and severity. This variation hampers the study of etiological differences and reduces the statistical power of analyses of associations to genetics, treatment outcomes, and complications. We address these issues through deep, fine-grained phenotypic stratification of a diabetes cohort. Text mining the electronic health records of 14,017 patients, we matched two controlled vocabularies (ICD-10 and a custom vocabulary developed at the clinical center Steno Diabetes Center Copenhagen) to clinical narratives spanning a 19 year period. The two matched vocabularies comprise over 20,000 medical terms describing symptoms, other diagnoses, and lifestyle factors. The cohort is genetically homogeneous (Caucasian diabetes patients from Denmark) so the resulting stratification is not driven by ethnic differences, but rather by inherently dissimilar progression patterns and lifestyle related risk factors. Using unsupervised Markov clustering, we defined 71 clusters of at least 50 individuals within the diabetes spectrum. The clusters display both distinct and shared longitudinal glycemic dysregulation patterns, temporal co-occurrences of comorbidities, and associations to single nucleotide polymorphisms in or near genes relevant for diabetes comorbidities.

Predicting dementia with routine care EMR data

Our aim is to develop a machine learning (ML) model that can predict dementia in a general patient population from multiple health care institutions one year and three years prior to the onset of the disease without any additional monitoring or screening. The purpose of the model is to automate the cost-effective, non-invasive, digital pre-screening of patients at risk for dementia. Towards this purpose, routine care data, which is widely available through Electronic Medical Record (EMR) systems is used as a data source. These data embody a rich knowledge and make related medical applications easy to deploy at scale in a cost-effective manner. Specifically, the model is trained by using structured and unstructured data from three EMR data sets: diagnosis, prescriptions, and medical notes. Each of these three data sets is used to construct an individual model along with a combined model which is derived by using all three data sets. Human-interpretable data processing and ML techniques are selected in order to facilitate adoption of the proposed model by health care providers from multiple institutions. The results show that the combined model is generalizable across multiple institutions and is able to predict dementia within one year of its onset with an accuracy of nearly 80% despite the fact that it was trained using routine care data. Moreover, the analysis of the models identified important predictors for dementia. Some of these predictors (e.g., age and hypertensive disorders) are already confirmed by the literature while others, especially the ones derived from the unstructured medical notes, require further clinical analysis.

Introducing Artificial Intelligence Training in Medical Education

Health care is evolving and with it the need to reform medical education. As the practice of medicine enters the age of artificial intelligence (AI), the use of data to improve clinical decision making will grow, pushing the need for skillful medicine-machine interaction. As the rate of medical knowledge grows, technologies such as AI are needed to enable health care professionals to effectively use this knowledge to practice medicine. Medical professionals need to be adequately trained in this new technology, its advantages to improve cost, quality, and access to health care, and its shortfalls such as transparency and liability. AI needs to be seamlessly integrated across different aspects of the curriculum. In this paper, we have addressed the state of medical education at present and have recommended a framework on how to evolve the medical education curriculum to include AI.

The role of artificial intelligence in medical imaging research

Without doubt, artificial intelligence (AI) is the most discussed topic today in medical imaging research, both in diagnostic and therapeutic. For diagnostic imaging alone, the number of publications on AI has increased from about 100-150 per year in 2007-2008 to 1000-1100 per year in 2017-2018. Researchers have applied AI to automatically recognizing complex patterns in imaging data and providing quantitative assessments of radiographic characteristics. In radiation oncology, AI has been applied on different image modalities that are used at different stages of the treatment. i.e. tumor delineation and treatment assessment. Radiomics, the extraction of a large number of image features from radiation images with a high-throughput approach, is one of the most popular research topics today in medical imaging research. AI is the essential boosting power of processing massive number of medical images and therefore uncovers disease characteristics that fail to be appreciated by the naked eyes. The objectives of this paper are to review the history of AI in medical imaging research, the current role, the challenges need to be resolved before AI can be adopted widely in the clinic, and the potential future.

Will Artificial Intelligence Translate Big Data Into Improved Medical Care or Be a Source of Confusing Intrusion? A Discussion Between a (Cautious) Physician Informatician and an (Optimistic) Medical Informatics Researcher

Artificial intelligence (AI), the computerized capability of doing tasks, which until recently was thought to be the exclusive domain of human intelligence, has demonstrated great strides in the past decade. The abilities to play games, provide piloting for an automobile, and respond to spoken language are remarkable successes. How are the challenges and opportunities of medicine different from these challenges and how can we best apply these data-driven techniques to patient care and outcomes? A New England Journal of Medicine paper published in 1980 suggested that more well-defined "specialized" tasks of medical care were more amenable to computer assistance, while the breadth of approach required for defining a problem and narrowing down the problem space was less so, and perhaps, unachievable. On the other hand, one can argue that the modern version of AI, which uses data-driven approaches, will be the most useful in tackling tasks such as outcome prediction that are often difficult for clinicians and patients. The ability today to collect large volumes of data about a single individual (eg, through a wearable device) and the accumulation of large datasets about multiple persons receiving medical care has the potential to apply to the care of individuals. As these techniques of analysis, enumeration, aggregation, and presentation are brought to bear in medicine, the question arises as to their utility and applicability in that domain. Early efforts in decision support were found to be helpful; as the systems proliferated, later experiences have shown difficulties such as alert fatigue and physician burnout becoming more prevalent. Will something similar arise from data-driven predictions? Will empowering patients by equipping them with information gained from data analysis help? Patients, providers, technology, and policymakers each have a role to play in the development and utilization of AI in medicine. Some of the challenges, opportunities, and tradeoffs implicit here are presented as a dialog between a clinician (SJN) and an informatician (QZT).

Distributed deep learning networks among institutions for medical imaging

Objective

Deep learning has become a promising approach for automated support for clinical diagnosis. When medical data samples are limited, collaboration among multiple institutions is necessary to achieve high algorithm performance. However, sharing patient data often has limitations due to technical, legal, or ethical concerns. In this study, we propose methods of distributing deep learning models as an attractive alternative to sharing patient data.

Methods

We simulate the distribution of deep learning models across 4 institutions using various training heuristics and compare the results with a deep learning model trained on centrally hosted patient data. The training heuristics investigated include ensembling single institution models, single weight transfer, and cyclical weight transfer. We evaluated these approaches for image classification in 3 independent image collections (retinal fundus photos, mammography, and ImageNet).

Results

We find that cyclical weight transfer resulted in a performance that was comparable to that of centrally hosted patient data. We also found that there is an improvement in the performance of cyclical weight transfer heuristic with a high frequency of weight transfer.

Conclusions

We show that distributing deep learning models is an effective alternative to sharing patient data. This finding has implications for any collaborative deep learning study.

Deep biomarkers of aging and longevity: from research to applications

Multiple recent advances in machine learning enabled computer systems to exceed human performance in many tasks including voice, text, and speech recognition and complex strategy games. Aging is a complex multifactorial process driven by and resulting in the many minute changes transpiring at every level of the human organism. Deep learning systems trained on the many measurable features changing in time can generalize and learn the many biological processes on the population and individual levels. The deep age predictors can help advance aging research by establishing causal relationships in non-linear systems. Deep aging clocks can be used for identification of novel therapeutic targets, evaluating the efficacy of the various interventions, data quality control, data economics, prediction of health trajectories, mortality, and many other applications. Here we present the current state of development of the deep aging clocks in the context of the pharmaceutical research and development and clinical applications.

Artificial intelligence in clinical and genomic diagnostics

Artificial intelligence (AI) is the development of computer systems that are able to perform tasks that normally require human intelligence. Advances in AI software and hardware, especially deep learning algorithms and the graphics processing units (GPUs) that power their training, have led to a recent and rapidly increasing interest in medical AI applications. In clinical diagnostics, AI-based computer vision approaches are poised to revolutionize image-based diagnostics, while other AI subtypes have begun to show similar promise in various diagnostic modalities. In some areas, such as clinical genomics, a specific type of AI algorithm known as deep learning is used to process large and complex genomic datasets. In this review, we first summarize the main classes of problems that AI systems are well suited to solve and describe the clinical diagnostic tasks that benefit from these solutions. Next, we focus on emerging methods for specific tasks in clinical genomics, including variant calling, genome annotation and variant classification, and phenotype-to-genotype correspondence. Finally, we end with a discussion on the future potential of AI in individualized medicine applications, especially for risk prediction in common complex diseases, and the challenges, limitations, and biases that must be carefully addressed for the successful deployment of AI in medical applications, particularly those utilizing human genetics and genomics data.

Liver disease screening based on densely connected deep neural networks

Liver disease is an important public health problem. Liver Function Tests (LFT) is the most achievable test for liver disease diagnosis. Most liver diseases are manifested as abnormal LFT. Liver disease screening by LFT data is helpful for computer aided diagnosis. In this paper, we propose a densely connected deep neural network (DenseDNN), on 13 most commonly used LFT indicators and demographic information of subjects for liver disease screening. The algorithm was tested on a dataset of 76,914 samples (more than 100 times of data than the previous datasets). The Area Under Curve (AUC) of DenseDNN is 0.8919, that of DNN is 0.8867, that of random forest is 0.8790, and that of logistic regression is 0.7974. The performance of deep learning models are significantly better than conventional methods. As for the deep learning methods, DenseDNN shows better performance than DNN.

Artificial Intelligence for Mental Health and Mental Illnesses: an Overview

PURPOSE OF REVIEW

Artificial intelligence (AI) technology holds both great promise to transform mental healthcare and potential pitfalls. This article provides an overview of AI and current applications in healthcare, a review of recent original research on AI specific to mental health, and a discussion of how AI can supplement clinical practice while considering its current limitations, areas needing additional research, and ethical implications regarding AI technology.

RECENT FINDINGS

We reviewed 28 studies of AI and mental health that used electronic health records (EHRs), mood rating scales, brain imaging data, novel monitoring systems (e.g., smartphone, video), and social media platforms to predict, classify, or subgroup mental health illnesses including depression, schizophrenia or other psychiatric illnesses, and suicide ideation and attempts. Collectively, these studies revealed high accuracies and provided excellent examples of AI's potential in mental healthcare, but most should be considered early proof-of-concept works demonstrating the potential of using machine learning (ML) algorithms to address mental health questions, and which types of algorithms yield the best performance. As AI techniques continue to be refined and improved, it will be possible to help mental health practitioners re-define mental illnesses more objectively than currently done in the DSM-5, identify these illnesses at an earlier or prodromal stage when interventions may be more effective, and personalize treatments based on an individual's unique characteristics. However, caution is necessary in order to avoid over-interpreting preliminary results, and more work is required to bridge the gap between AI in mental health research and clinical care.

Deep neural networks in psychiatry

Machine and deep learning methods, today's core of artificial intelligence, have been applied with increasing success and impact in many commercial and research settings. They are powerful tools for large scale data analysis, prediction and classification, especially in very data-rich environments ("big data"), and have started to find their way into medical applications. Here we will first give an overview of machine learning methods, with a focus on deep and recurrent neural networks, their relation to statistics, and the core principles behind them. We will then discuss and review directions along which (deep) neural networks can be, or already have been, applied in the context of psychiatry, and will try to delineate their future potential in this area. We will also comment on an emerging area that so far has been much less well explored: by embedding semantically interpretable computational models of brain dynamics or behavior into a statistical machine learning context, insights into dysfunction beyond mere prediction and classification may be gained. Especially this marriage of computational models with statistical inference may offer insights into neural and behavioral mechanisms that could open completely novel avenues for psychiatric treatment.

Toward a clinical text encoder: pretraining for clinical natural language processing with applications to substance misuse

OBJECTIVE

Our objective is to develop algorithms for encoding clinical text into representations that can be used for a variety of phenotyping tasks.

MATERIALS AND METHODS

Obtaining large datasets to take advantage of highly expressive deep learning methods is difficult in clinical natural language processing (NLP). We address this difficulty by pretraining a clinical text encoder on billing code data, which is typically available in abundance. We explore several neural encoder architectures and deploy the text representations obtained from these encoders in the context of clinical text classification tasks. While our ultimate goal is learning a universal clinical text encoder, we also experiment with training a phenotype-specific encoder. A universal encoder would be more practical, but a phenotype-specific encoder could perform better for a specific task.

RESULTS

We successfully train several clinical text encoders, establish a new state-of-the-art on comorbidity data, and observe good performance gains on substance misuse data.

DISCUSSION

We find that pretraining using billing codes is a promising research direction. The representations generated by this type of pretraining have universal properties, as they are highly beneficial for many phenotyping tasks. Phenotype-specific pretraining is a viable route for trading the generality of the pretrained encoder for better performance on a specific phenotyping task.

CONCLUSIONS

We successfully applied our approach to many phenotyping tasks. We conclude by discussing potential limitations of our approach.

Current Techniques for Diabetes Prediction: Review and Case Study

Diabetes is one of the most common diseases worldwide. Many Machine Learning (ML) techniques have been utilized in predicting diabetes in the last couple of years. The increasing complexity of this problem has inspired researchers to explore the robust set of Deep Learning (DL) algorithms. The highest accuracy achieved so far was 95.1% by a combined model CNN-LSTM. Even though numerous ML algorithms were used in solving this problem, there are a set of classifiers that are rarely used or even not used at all in this problem, so it is of interest to determine the performance of these classifiers in predicting diabetes. Moreover, there is no recent survey that has reviewed and compared the performance of all the proposed ML and DL techniques in addition to combined models. This article surveyed all the ML and DL techniques-based diabetes predictions published in the last six years. In addition, one study was developed that aimed to implement those rarely and not used ML classifiers on the Pima Indian Dataset to analyze their performance. The classifiers obtained an accuracy of 68%–74%. The recommendation is to use these classifiers in diabetes prediction and enhance them by developing combined models.