Machine Learning and Electronic Health Records: A Paradigm Shift

Machine learning to predict mortality for aneurysmal subarachnoid hemorrhage (aSAH) using a large nationwide EHR database

Aneurysmal subarachnoid hemorrhage (aSAH) develops quickly once it occurs and threatens the life of patients. We aimed to use machine learning to predict mortality for SAH patients at an early stage which can help doctors make clinical decisions. In our study, we applied different machine learning methods to an aSAH cohort extracted from a national EHR database, the Cerner Health Facts EHR database (2000-2018). The outcome of interest was in-hospital mortality, as either passing away while still in the hospital or being discharged to hospice care. Machine learning-based models were primarily evaluated by the area under the receiver operating characteristic curve (AUC). The population size of the SAH cohort was 6728. The machine learning methods achieved an average of AUCs of 0.805 for predicting mortality with only the initial 24 hours' EHR data. Without losing the prediction power, we used the logistic regression to identify 42 risk factors, -examples include age and serum glucose-that exhibit a significant correlation with the mortality of aSAH patients. Our study illustrates the potential of utilizing machine learning techniques as a practical prognostic tool for predicting aSAH mortality at the bedside.

In-hospital fall prediction using machine learning algorithms and the Morse fall scale in patients with acute stroke: a nested case-control study

BACKGROUND

Falls are one of the most common accidents in medical institutions, which can threaten the safety of inpatients and negatively affect their prognosis. Herein, we developed a machine learning (ML) model for fall prediction in patients with acute stroke and compared its accuracy with that of the existing fall risk prediction tool, the Morse Fall Scale (MFS).

METHODS

This is a retrospective nested case-control study. The initial sample size was 8462 admitted to a single cerebrovascular specialty hospital with acute stroke. A total of 156 fall events occurred, and each fall case was randomly matched with six control cases. Six ML algorithms were used, namely, regularized logistic regression, support vector machine, naïve Bayes (NB), k-nearest neighbors, random forest, and extreme-gradient boosting (XGB).

RESULTS

We included 156 in the fall group and 934 in the non-fall group. The mean ages of the fall and non-fall groups were 68.3 (± 12.2) and 65.3 (± 12.9) years old, respectively. The MFS total score was significantly higher in the fall group (54.3 ± 18.3) than in the non-fall group (37.7 ± 14.7). The area under the receiver operating curve (AUROC) of the MFS in predicting falls was 0.76 (0.73-0.79). XGB had the highest AUROC of 0.85 (0.78-0.92), and XGB and NB had the highest F1 score of 0.44.

CONCLUSIONS

The AUROC values of all of ML algorithms were similar to those of the MFS in predicting fall risk in patients with acute stroke, allowing for accurate and efficient fall screening.

Mental Health Information Reporting Assistant (MHIRA)-an open-source software facilitating evidence-based assessment for clinical services

Evidence-based assessment (EBA) in mental health is a critical aspect of improving patient outcomes and addressing the gaps in mental health care. EBA involves the use of psychometric instruments to gather data that can inform clinical decision-making, inform policymakers, and serve as a basis for research and quality management. Despite its potential, EBA is often hindered by barriers such as workload and cost, leading to its underutilization. Regarding low- and middle-income countries (LMIC), the implementation of EBA is recognized as a key strategy to address and close the prevalent mental health treatment gap.To simplify the application of EBA including in LMIC, an international team of researchers and practitioners from Tanzania, Kosovo, Chile, and Switzerland developed the Mental Health Information Reporting Assistant (MHIRA). MHIRA is an open-source electronic health record that streamlines EBA by digitising psychometric instruments and organising patient data in a user-friendly manner. It provides immediate and convenient reports to inform clinical decision-making.The current article provides a comprehensive overview of the features and technical details of MHIRA, as well as insights from four implementation scenarios. The experience gained during the implementations as well as the user-feedback suggests that MHIRA has the potential to be successfully implemented in a variety of clinical contexts and simplify the use of EBA. However, further research is necessary to establish its potential to sustainably transform healthcare services and impact patient outcomes.In conclusion, MHIRA represents an important step in promoting the widespread adoption of EBA in mental health. It offers a promising solution to the barriers that have limited the use of EBA in the past and holds the potential to improve patient outcomes and support the ongoing efforts to address gaps in mental health care.

Disclosure control of machine learning models from trusted research environments (TRE): New challenges and opportunities

Introduction

Artificial intelligence (AI) applications in healthcare and medicine have increased in recent years. To enable access to personal data, Trusted Research Environments (TREs) (otherwise known as Safe Havens) provide safe and secure environments in which researchers can access sensitive personal data and develop AI (in particular machine learning (ML)) models. However, currently few TREs support the training of ML models in part due to a gap in the practical decision-making guidance for TREs in handling model disclosure. Specifically, the training of ML models creates a need to disclose new types of outputs from TREs. Although TREs have clear policies for the disclosure of statistical outputs, the extent to which trained models can leak personal training data once released is not well understood.

Background

We review, for a general audience, different types of ML models and their applicability within healthcare. We explain the outputs from training a ML model and how trained ML models can be vulnerable to external attacks to discover personal data encoded within the model.

Risks

We present the challenges for disclosure control of trained ML models in the context of training and exporting models from TREs. We provide insights and analyse methods that could be introduced within TREs to mitigate the risk of privacy breaches when disclosing trained models.

Discussion

Although specific guidelines and policies exist for statistical disclosure controls in TREs, they do not satisfactorily address these new types of output requests; i.e., trained ML models. There is significant potential for new interdisciplinary research opportunities in developing and adapting policies and tools for safely disclosing ML outputs from TREs.

A Framework for Automatic Clustering of EHR Messages Using a Spatial Clustering Approach

Although Health Level Seven (HL 7) message standards (v2, v3, Clinical Document Architecture (CDA)) have been commonly adopted, there are still issues associated with them, especially the semantic interoperability issues and lack of support for smart devices (e.g., smartphones, fitness trackers, and smartwatches), etc. In addition, healthcare organizations in many countries are still using proprietary electronic health record (EHR) message formats, making it challenging to convert to other data formats-particularly the latest HL7 Fast Health Interoperability Resources (FHIR) data standard. The FHIR is based on modern web technologies such as HTTP, XML, and JSON and would be capable of overcoming the shortcomings of the previous standards and supporting modern smart devices. Therefore, the FHIR standard could help the healthcare industry to avail the latest technologies benefits and improve data interoperability. The data representation and mapping from the legacy data standards (i.e., HL7 v2 and EHR) to the FHIR is necessary for the healthcare sector. However, direct data mapping or conversion from the traditional data standards to the FHIR data standard is challenging because of the nature and formats of the data. Therefore, in this article, we propose a framework that aims to convert proprietary EHR messages into the HL7 v2 format and apply an unsupervised clustering approach using the DBSCAN (density-based spatial clustering of applications with noise) algorithm to automatically group a variety of these HL7 v2 messages regardless of their semantic origins. The proposed framework's implementation lays the groundwork to provide a generic mapping model with multi-point and multi-format data conversion input into the FHIR. Our experimental results show the proposed framework's ability to automatically cluster various HL7 v2 message formats and provide analytic insight behind them.

Machine Learning Models for the Prediction of Renal Failure in Chronic Kidney Disease: A Retrospective Cohort Study

This study assessed the feasibility of five separate machine learning (ML) classifiers for predicting disease progression in patients with pre-dialysis chronic kidney disease (CKD). The study enrolled 858 patients with CKD treated at a veteran’s hospital in Taiwan. After classification into early and advanced stages, patient demographics and laboratory data were processed and used to predict progression to renal failure and important features for optimal prediction were identified. The random forest (RF) classifier with synthetic minority over-sampling technique (SMOTE) had the best predictive performances among patients with early-stage CKD who progressed within 3 and 5 years and among patients with advanced-stage CKD who progressed within 1 and 3 years. Important features identified for predicting progression from early- and advanced-stage CKD were urine creatinine and serum creatinine levels, respectively. The RF classifier demonstrated the optimal performance, with an area under the receiver operating characteristic curve values of 0.96 for predicting progression within 5 years in patients with early-stage CKD and 0.97 for predicting progression within 1 year in patients with advanced-stage CKD. The proposed method resulted in the optimal prediction of CKD progression, especially within 1 year of advanced-stage CKD. These results will be useful for predicting prognosis among patients with CKD.

Ten simple rules for organizing a special session at a scientific conference

Special sessions are important parts of scientific meetings and conferences: They gather together researchers and students interested in a specific topic and can strongly contribute to the success of the conference itself. Moreover, they can be the first step for trainees and students to the organization of a scientific event. Organizing a special session, however, can be uneasy for beginners and students. Here, we provide ten simple rules to follow to organize a special session at a scientific conference.

Predicting the Need for Therapeutic Intervention and Mortality in Acute Pancreatitis: A Two-Center International Study Using Machine Learning

Background: Current approaches to predicting intervention needs and mortality have reached 65–85% accuracy, which falls below clinical decision-making requirements in patients with acute pancreatitis (AP). We aimed to accurately predict therapeutic intervention needs and mortality on admission, in AP patients, using machine learning (ML). Methods: Data were obtained from three databases of patients admitted with AP: one retrospective (Chengdu) and two prospective (Liverpool and Chengdu) databases. Intervention and mortality differences, as well as potential predictors, were investigated. Univariate analysis was conducted, followed by a random forest ML algorithm used in multivariate analysis, to identify predictors. The ML performance matrix was applied to evaluate the model’s performance. Results: Three datasets of 2846 patients included 25 potential clinical predictors in the univariate analysis. The top ten identified predictors were obtained by ML models, for predicting interventions and mortality, from the training dataset. The prediction of interventions includes death in non-intervention patients, validated with high accuracy (96%/98%), the area under the receiver-operating-characteristic curve (0.90/0.98), and positive likelihood ratios (22.3/69.8), respectively. The post-test probabilities in the test set were 55.4% and 71.6%, respectively, which were considerably superior to existing prognostic scores. The ML model, for predicting mortality in intervention patients, performed better or equally with prognostic scores. Conclusions: ML, using admission clinical predictors, can accurately predict therapeutic interventions and mortality in patients with AP.

Towards Validating the Effectiveness of Obstructive Sleep Apnea Classification from Electronic Health Records Using Machine Learning

Obstructive sleep apnea (OSA) is a common, chronic, sleep-related breathing disorder characterized by partial or complete airway obstruction in sleep. The gold standard diagnosis method is polysomnography, which estimates disease severity through the Apnea-Hypopnea Index (AHI). However, this is expensive and not widely accessible to the public. For effective screening, this work implements machine learning algorithms for classification of OSA. The model is trained with routinely acquired clinical data of 1479 records from the Wisconsin Sleep Cohort dataset. Extracted features from the electronic health records include patient demographics, laboratory blood reports, physical measurements, habitual sleep history, comorbidities, and general health questionnaire scores. For distinguishing between OSA and non-OSA patients, feature selection methods reveal the primary important predictors as waist-to-height ratio, waist circumference, neck circumference, body-mass index, lipid accumulation product, excessive daytime sleepiness, daily snoring frequency and snoring volume. Optimal hyperparameters were selected using a hybrid tuning method consisting of Bayesian Optimization and Genetic Algorithms through a five-fold cross-validation strategy. Support vector machines achieved the highest evaluation scores with accuracy: 68.06%, sensitivity: 88.76%, specificity: 40.74%, F1-score: 75.96%, PPV: 66.36% and NPV: 73.33%. We conclude that routine clinical data can be useful in prioritization of patient referral for further sleep studies.

Leveraging the potential of machine learning for assessing vascular ageing: state-of-the-art and future research

Vascular ageing biomarkers have been found to be predictive of cardiovascular risk independently of classical risk factors, yet are not widely used in clinical practice. In this review, we present two basic approaches for using machine learning (ML) to assess vascular age: parameter estimation and risk classification. We then summarize their role in developing new techniques to assess vascular ageing quickly and accurately. We discuss the methods used to validate ML-based markers, the evidence for their clinical utility, and key directions for future research. The review is complemented by case studies of the use of ML in vascular age assessment which can be replicated using freely available data and code.

Machine Learning Prediction Models for Chronic Kidney Disease Using National Health Insurance Claim Data in Taiwan

Chronic kidney disease (CKD) represents a heavy burden on the healthcare system because of the increasing number of patients, high risk of progression to end-stage renal disease, and poor prognosis of morbidity and mortality. The aim of this study is to develop a machine-learning model that uses the comorbidity and medication data obtained from Taiwan's National Health Insurance Research Database to forecast the occurrence of CKD within the next 6 or 12 months before its onset, and hence its prevalence in the population. A total of 18,000 people with CKD and 72,000 people without CKD diagnosis were selected using propensity score matching. Their demographic, medication and comorbidity data from their respective two-year observation period were used to build a predictive model. Among the approaches investigated, the Convolutional Neural Networks (CNN) model performed best with a test set AUROC of 0.957 and 0.954 for the 6-month and 12-month predictions, respectively. The most prominent predictors in the tree-based models were identified, including diabetes mellitus, age, gout, and medications such as sulfonamides and angiotensins. The model proposed in this study could be a useful tool for policymakers in predicting the trends of CKD in the population. The models can allow close monitoring of people at risk, early detection of CKD, better allocation of resources, and patient-centric management.

Natural language processing and machine learning of electronic health records for prediction of first-time suicide attempts

OBJECTIVE

Limited research exists in predicting first-time suicide attempts that account for two-thirds of suicide decedents. We aimed to predict first-time suicide attempts using a large data-driven approach that applies natural language processing (NLP) and machine learning (ML) to unstructured (narrative) clinical notes and structured electronic health record (EHR) data.

METHODS

This case-control study included patients aged 10-75 years who were seen between 2007 and 2016 from emergency departments and inpatient units. Cases were first-time suicide attempts from coded diagnosis; controls were randomly selected without suicide attempts regardless of demographics, following a ratio of nine controls per case. Four data-driven ML models were evaluated using 2-year historical EHR data prior to suicide attempt or control index visits, with prediction windows from 7 to 730 days. Patients without any historical notes were excluded. Model evaluation on accuracy and robustness was performed on a blind dataset (30% cohort).

RESULTS

The study cohort included 45 238 patients (5099 cases, 40 139 controls) comprising 54 651 variables from 5.7 million structured records and 798 665 notes. Using both unstructured and structured data resulted in significantly greater accuracy compared to structured data alone (area-under-the-curve [AUC]: 0.932 vs. 0.901 P < .001). The best-predicting model utilized 1726 variables with AUC = 0.932 (95% CI, 0.922-0.941). The model was robust across multiple prediction windows and subgroups by demographics, points of historical most recent clinical contact, and depression diagnosis history.

CONCLUSIONS

Our large data-driven approach using both structured and unstructured EHR data demonstrated accurate and robust first-time suicide attempt prediction, and has the potential to be deployed across various populations and clinical settings.

Robust data-driven identification of risk factors and their interactions: A simulation and a study of parental and demographic risk factors for schizophrenia

OBJECTIVES

Few interactions between risk factors for schizophrenia have been replicated, but fitting all such interactions is difficult due to high-dimensionality. Our aims are to examine significant main and interaction effects for schizophrenia and the performance of our approach using simulated data.

METHODS

We apply the machine learning technique elastic net to a high-dimensional logistic regression model to produce a sparse set of predictors, and then assess the significance of odds ratios (OR) with Bonferroni-corrected p-values and confidence intervals (CI). We introduce a simulation model that resembles a Finnish nested case-control study of schizophrenia which uses national registers to identify cases (n = 1,468) and controls (n = 2,975). The predictors include nine sociodemographic factors and all interactions (31 predictors).

RESULTS

In the simulation, interactions with OR = 3 and prevalence = 4% were identified with <5% false positive rate and ≥80% power. None of the studied interactions were significantly associated with schizophrenia, but main effects of parental psychosis (OR = 5.2, CI 2.9-9.7; p < .001), urbanicity (1.3, 1.1-1.7; p = .001), and paternal age ≥35 (1.3, 1.004-1.6; p = .04) were significant.

CONCLUSIONS

We have provided an analytic pipeline for data-driven identification of main and interaction effects in case-control data. We identified highly replicated main effects for schizophrenia, but no interactions.

What is Interpretability?

We argue that artificial networks are explainable and offer a novel theory of interpretability. Two sets of conceptual questions are prominent in theoretical engagements with artificial neural networks, especially in the context of medical artificial intelligence: (1) Are networks explainable, and if so, what does it mean to explain the output of a network? And (2) what does it mean for a network to be interpretable? We argue that accounts of "explanation" tailored specifically to neural networks have ineffectively reinvented the wheel. In response to (1), we show how four familiar accounts of explanation apply to neural networks as they would to any scientific phenomenon. We diagnose the confusion about explaining neural networks within the machine learning literature as an equivocation on "explainability," "understandability" and "interpretability." To remedy this, we distinguish between these notions, and answer (2) by offering a theory and typology of interpretation in machine learning. Interpretation is something one does to an explanation with the aim of producing another, more understandable, explanation. As with explanation, there are various concepts and methods involved in interpretation: Total or Partial, Global or Local, and Approximative or Isomorphic. Our account of "interpretability" is consistent with uses in the machine learning literature, in keeping with the philosophy of explanation and understanding, and pays special attention to medical artificial intelligence systems.

Genome-Wide Association Study of Suicide Death and Polygenic Prediction of Clinical Antecedents

OBJECTIVE

Death by suicide is a highly preventable yet growing worldwide health crisis. To date, there has been a lack of adequately powered genomic studies of suicide, with no sizable suicide death cohorts available for analysis. To address this limitation, the authors conducted the first comprehensive genomic analysis of suicide death using previously unpublished genotype data from a large population-ascertained cohort.

METHODS

The analysis sample comprised 3,413 population-ascertained case subjects of European ancestry and 14,810 ancestrally matched control subjects. Analytical methods included principal component analysis for ancestral matching and adjusting for population stratification, linear mixed model genome-wide association testing (conditional on genetic-relatedness matrix), gene and gene set-enrichment testing, and polygenic score analyses, as well as single-nucleotide polymorphism (SNP) heritability and genetic correlation estimation using linkage disequilibrium score regression.

RESULTS

Genome-wide association analysis identified two genome-wide significant loci (involving six SNPs: rs34399104, rs35518298, rs34053895, rs66828456, rs35502061, and rs35256367). Gene-based analyses implicated 22 genes on chromosomes 13, 15, 16, 17, and 19 (q<0.05). Suicide death heritability was estimated at an h2SNP value of 0.25 (SE=0.04) and a value of 0.16 (SE=0.02) when converted to a liability scale. Notably, suicide polygenic scores were significantly predictive across training and test sets. Polygenic scores for several other psychiatric disorders and psychological traits were also predictive, particularly scores for behavioral disinhibition and major depressive disorder.

CONCLUSIONS

Multiple genome-wide significant loci and genes were identified and polygenic score prediction of suicide death case-control status was demonstrated, adjusting for ancestry, in independent training and test sets. Additionally, the suicide death sample was found to have increased genetic risk for behavioral disinhibition, major depressive disorder, depressive symptoms, autism spectrum disorder, psychosis, and alcohol use disorder compared with the control sample.

Identifying the Medical Lethality of Suicide Attempts Using Network Analysis and Deep Learning: Nationwide Study

BACKGROUND

Suicide is one of the leading causes of death among young and middle-aged people. However, little is understood about the behaviors leading up to actual suicide attempts and whether these behaviors are specific to the nature of suicide attempts.

OBJECTIVE

The goal of this study was to examine the clusters of behaviors antecedent to suicide attempts to determine if they could be used to assess the potential lethality of the attempt. To accomplish this goal, we developed a deep learning model using the relationships among behaviors antecedent to suicide attempts and the attempts themselves.

METHODS

This study used data from the Korea National Suicide Survey. We identified 1112 individuals who attempted suicide and completed a psychiatric evaluation in the emergency room. The 15-item Beck Suicide Intent Scale (SIS) was used for assessing antecedent behaviors, and the medical outcomes of the suicide attempts were measured by assessing lethality with the Columbia Suicide Severity Rating Scale (C-SSRS; lethal suicide attempt >3 and nonlethal attempt ≤3).

RESULTS

Using scores from the SIS, individuals who had lethal and nonlethal attempts comprised two different network nodes with the edges representing the relationships among nodes. Among the antecedent behaviors, the conception of a method's lethality predicted suicidal behaviors with severe medical outcomes. The vectorized relationship values among the elements of antecedent behaviors in our deep learning model (E-GONet) increased performances, such as F1 and area under the precision-recall gain curve (AUPRG), for identifying lethal attempts (up to 3% for F1 and 32% for AUPRG), as compared with other models (mean F1: 0.81 for E-GONet, 0.78 for linear regression, and 0.80 for random forest; mean AUPRG: 0.73 for E-GONet, 0.41 for linear regression, and 0.69 for random forest).

CONCLUSIONS

The relationships among behaviors antecedent to suicide attempts can be used to understand the suicidal intent of individuals and help identify the lethality of potential suicide attempts. Such a model may be useful in prioritizing cases for preventive intervention.

Risk and Protective Factors of Current Opioid Use Among Youth Living on or Near American Indian Reservations: An Application of Machine Learning

Opioid use among youth, particularly among American Indian (AI) youth, is rising, resulting in a large number of accidental overdoses and deaths. In order to develop effective prevention strategies, we need to use exploratory data analysis to identify previously unknown predictors of opioid use among youth living on or near reservations. The present study is an application of Machine Learning, a type of exploratory data analysis, to the Our Youth, Our Future epidemiological survey (N = 6482) to determine salient risk and protective factors for past 30-day opioid use. The Machine Learning algorithm identified 11 salient risk and protective factors. Importantly, highest risk was conferred for those reporting recent cocaine use, having ever tried a narcotic other than heroin, and identifying as American Indian. Protective factors included never having tried opioids other than heroin, infrequent binge drinking, having fewer friends pressuring you to use illicit drugs, initiating alcohol use at a later age, and being older. This model explained 61% of the variance in the training sample and, on average, 24% of the variance in the bootstrapped samples. Taken together, this model identifies known predictors of 30-day opioid use, for example, recent substance use, as well as unknown predictors including being AI, Snapchat use, and peer encouragement for use. Notably, recent cocaine use was a more salient predictor of recent opioid use than lifetime opioid use.

The emerging landscape of health research based on biobanks linked to electronic health records: Existing resources, statistical challenges, and potential opportunities

Biobanks linked to electronic health records provide rich resources for health-related research. With improvements in administrative and informatics infrastructure, the availability and utility of data from biobanks have dramatically increased. In this paper, we first aim to characterize the current landscape of available biobanks and to describe specific biobanks, including their place of origin, size, and data types. The development and accessibility of large-scale biorepositories provide the opportunity to accelerate agnostic searches, expedite discoveries, and conduct hypothesis-generating studies of disease-treatment, disease-exposure, and disease-gene associations. Rather than designing and implementing a single study focused on a few targeted hypotheses, researchers can potentially use biobanks' existing resources to answer an expanded selection of exploratory questions as quickly as they can analyze them. However, there are many obvious and subtle challenges with the design and analysis of biobank-based studies. Our second aim is to discuss statistical issues related to biobank research such as study design, sampling strategy, phenotype identification, and missing data. We focus our discussion on biobanks that are linked to electronic health records. Some of the analytic issues are illustrated using data from the Michigan Genomics Initiative and UK Biobank, two biobanks with two different recruitment mechanisms. We summarize the current body of literature for addressing these challenges and discuss some standing open problems. This work complements and extends recent reviews about biobank-based research and serves as a resource catalog with analytical and practical guidance for statisticians, epidemiologists, and other medical researchers pursuing research using biobanks.

Unlocking the Potential of Electronic Health Records for Health Research

Electronic health records (EHRs), originally designed to facilitate health care delivery, are becoming a valuable data source for health research. EHR systems have two components, both of which have various components, and points of data entry, management, and analysis. The “front end” refers to where the data are entered, primarily by healthcare workers (e.g. physicians and nurses). The second component of EHR systems is the electronic data warehouse, or “back-end,” where the data are stored in a relational database. EHR data elements can be of many types, which can be categorized as structured, unstructured free-text, and imaging data. The Sunrise Clinical Manager (SCM) EHR is one example of an inpatient EHR system, which covers the city of Calgary (Alberta, Canada). This system, under the management of Alberta Health Services, is now being explored for research use. The purpose of the present paper is to describe the SCM EHR for research purposes, showing how this generalizes to EHRs in general. We further discuss advantages, challenges (e.g. potential bias and data quality issues), analytical capacities, and requirements associated with using EHRs in a health research context.

Beyond the Randomized Clinical Trial: Innovative Data Science to Close the Pediatric Evidence Gap

Despite the application of advanced statistical and pharmacometric approaches to pediatric trial data, a large pediatric evidence gap still remains. Here, we discuss how to collect more data from children by using real-world data from electronic health records, mobile applications, wearables, and social media. The large datasets collected with these approaches enable and may demand the use of artificial intelligence and machine learning to allow the data to be analyzed for decision making. Applications of this approach are presented, which include the prediction of future clinical complications, medical image analysis, identification of new pediatric end points and biomarkers, the prediction of treatment nonresponders, and the prediction of placebo-responders for trial enrichment. Finally, we discuss how to bring machine learning from science to pediatric clinical practice. We conclude that advantage should be taken of the current opportunities offered by innovations in data science and machine learning to close the pediatric evidence gap.

Interpreting patient-Specific risk prediction using contextual decomposition of BiLSTMs: application to children with asthma

BACKGROUND

Predictive modeling with longitudinal electronic health record (EHR) data offers great promise for accelerating personalized medicine and better informs clinical decision-making. Recently, deep learning models have achieved state-of-the-art performance for many healthcare prediction tasks. However, deep models lack interpretability, which is integral to successful decision-making and can lead to better patient care. In this paper, we build upon the contextual decomposition (CD) method, an algorithm for producing importance scores from long short-term memory networks (LSTMs). We extend the method to bidirectional LSTMs (BiLSTMs) and use it in the context of predicting future clinical outcomes using patients' EHR historical visits.

METHODS

We use a real EHR dataset comprising 11071 patients, to evaluate and compare CD interpretations from LSTM and BiLSTM models. First, we train LSTM and BiLSTM models for the task of predicting which pre-school children with respiratory system-related complications will have asthma at school-age. After that, we conduct quantitative and qualitative analysis to evaluate the CD interpretations produced by the contextual decomposition of the trained models. In addition, we develop an interactive visualization to demonstrate the utility of CD scores in explaining predicted outcomes.

RESULTS

Our experimental evaluation demonstrate that whenever a clear visit-level pattern exists, the models learn that pattern and the contextual decomposition can appropriately attribute the prediction to the correct pattern. In addition, the results confirm that the CD scores agree to a large extent with the importance scores generated using logistic regression coefficients. Our main insight was that rather than interpreting the attribution of individual visits to the predicted outcome, we could instead attribute a model's prediction to a group of visits.

CONCLUSION

We presented a quantitative and qualitative evidence that CD interpretations can explain patient-specific predictions using CD attributions of individual visits or a group of visits.

Developing a FHIR-based EHR phenotyping framework: A case study for identification of patients with obesity and multiple comorbidities from discharge summaries

BACKGROUND

Standards-based clinical data normalization has become a key component of effective data integration and accurate phenotyping for secondary use of electronic healthcare records (EHR) data. HL7 Fast Healthcare Interoperability Resources (FHIR) is an emerging clinical data standard for exchanging electronic healthcare data and has been used in modeling and integrating both structured and unstructured EHR data for a variety of clinical research applications. The overall objective of this study is to develop and evaluate a FHIR-based EHR phenotyping framework for identification of patients with obesity and its multiple comorbidities from semi-structured discharge summaries leveraging a FHIR-based clinical data normalization pipeline (known as NLP2FHIR).

METHODS

We implemented a multi-class and multi-label classification system based on the i2b2 Obesity Challenge task to evaluate the FHIR-based EHR phenotyping framework. Two core parts of the framework are: (a) the conversion of discharge summaries into corresponding FHIR resources - Composition, Condition, MedicationStatement, Procedure and FamilyMemberHistory using the NLP2FHIR pipeline, and (b) the implementation of four machine learning algorithms (logistic regression, support vector machine, decision tree, and random forest) to train classifiers to predict disease state of obesity and 15 comorbidities using features extracted from standard FHIR resources and terminology expansions. We used the macro- and micro-averaged precision (P), recall (R), and F1 score (F1) measures to evaluate the classifier performance. We validated the framework using a second obesity dataset extracted from the MIMIC-III database.

RESULTS

Using the NLP2FHIR pipeline, 1237 clinical discharge summaries from the 2008 i2b2 obesity challenge dataset were represented as the instances of the FHIR Composition resource consisting of 5677 records with 16 unique section types. After the NLP processing and FHIR modeling, a set of 244,438 FHIR clinical resource instances were generated. As the results of the four machine learning classifiers, the random forest algorithm performed the best with F1-micro(0.9466)/F1-macro(0.7887) and F1-micro(0.9536)/F1-macro(0.6524) for intuitive classification (reflecting medical professionals' judgments) and textual classification (reflecting the judgments based on explicitly reported information of diseases), respectively. The MIMIC-III obesity dataset was successfully integrated for prediction with minimal configuration of the NLP2FHIR pipeline and machine learning models.

CONCLUSIONS

The study demonstrated that the FHIR-based EHR phenotyping approach could effectively identify the state of obesity and multiple comorbidities using semi-structured discharge summaries. Our FHIR-based phenotyping approach is a first concrete step towards improving the data aspect of phenotyping portability across EHR systems and enhancing interpretability of the machine learning-based phenotyping algorithms.

Predicting Suicide Attempts and Suicide Deaths Following Outpatient Visits Using Electronic Health Records

OBJECTIVE

The authors sought to develop and validate models using electronic health records to predict suicide attempt and suicide death following an outpatient visit.

METHOD

Across seven health systems, 2,960,929 patients age 13 or older (mean age, 46 years; 62% female) made 10,275,853 specialty mental health visits and 9,685,206 primary care visits with mental health diagnoses between Jan. 1, 2009, and June 30, 2015. Health system records and state death certificate data identified suicide attempts (N=24,133) and suicide deaths (N=1,240) over 90 days following each visit. Potential predictors included 313 demographic and clinical characteristics extracted from records for up to 5 years before each visit: prior suicide attempts, mental health and substance use diagnoses, medical diagnoses, psychiatric medications dispensed, inpatient or emergency department care, and routinely administered depression questionnaires. Logistic regression models predicting suicide attempt and death were developed using penalized LASSO (least absolute shrinkage and selection operator) variable selection in a random sample of 65% of the visits and validated in the remaining 35%.

RESULTS

Mental health specialty visits with risk scores in the top 5% accounted for 43% of subsequent suicide attempts and 48% of suicide deaths. Of patients scoring in the top 5%, 5.4% attempted suicide and 0.26% died by suicide within 90 days. C-statistics (equivalent to area under the curve) for prediction of suicide attempt and suicide death were 0.851 (95% CI=0.848, 0.853) and 0.861 (95% CI=0.848, 0.875), respectively. Primary care visits with scores in the top 5% accounted for 48% of subsequent suicide attempts and 43% of suicide deaths. C-statistics for prediction of suicide attempt and suicide death were 0.853 (95% CI=0.849, 0.857) and 0.833 (95% CI=0.813, 0.853), respectively.

CONCLUSIONS

Prediction models incorporating both health record data and responses to self-report questionnaires substantially outperform existing suicide risk prediction tools.

Big Brain Data: On the Responsible Use of Brain Data from Clinical and Consumer-Directed Neurotechnological Devices

The focus of this paper are the ethical, legal and social challenges for ensuring the responsible use of “big brain data”—the recording, collection and analysis of individuals’ brain data on a large scale with clinical and consumer-directed neurotechnological devices. First, I highlight the benefits of big data and machine learning analytics in neuroscience for basic and translational research. Then, I describe some of the technological, social and psychological barriers for securing brain data from unwarranted access. In this context, I then examine ways in which safeguards at the hardware and software level, as well as increasing “data literacy” in society, may enhance the security of neurotechnological devices and protect the privacy of personal brain data. Regarding ethical and legal ramifications of big brain data, I first discuss effects on the autonomy, the sense of agency and authenticity, as well as the self that may result from the interaction between users and intelligent, particularly closed-loop, neurotechnological devices. I then discuss the impact of the “datafication” in basic and clinical neuroscience research on the just distribution of resources and access to these transformative technologies. In the legal realm, I examine possible legal consequences that arises from the increasing abilities to decode brain states and their corresponding subjective phenomenological experiences on the hitherto inaccessible privacy of these information. Finally, I discuss the implications of big brain data for national and international regulatory policies and models of good data governance.

Leveraging psychiatric and medical genetics to understand comorbid depression and obesity

Precision medicine in psychiatry is on the rise, and depression and obesity - two highly prevalent, comorbid and well-characterised phenotypes - are optimal targets for the approach. Add the bedrock susceptibility gene, FTO, and Riviera et al have identified a constellation of factors that could enhance clinical treatment of both disorders.