A Reproducible Model Based on Clinical Text for Predicting Suicidal Behavior.

Suicide risk models are critical for prioritizing patients for intervention. We demonstrate a reproducible approach for training text classifiers to identify patients at risk. The models were effective in phenotyping suicidal behavior (F1=.94) and moderately effective in predicting future events (F1=.63).

Overcoming recruitment hurdles in clinical trials: An investigation of remote consenting modalities and their impact on workload, workflow, and usability.

Travel constraints can hinder the recruitment of eligible research participants in clinical trials, causing research timeline extensions, added costs, underpowered results, and early termination. Remote consenting can help solve these issues by allowing researchers and potential participants to connect remotely. This controlled experimental study investigates the effect of remote consenting on workload, workflow, usability, and barriers and facilitators to its implementation. Using a mixed experimental design, simulated consenting sessions were conducted with three different modalities (remote paper consent, eConsent, and teleconsent) with 23 researchers and 93 research participants. Each session involved a single researcher who experienced all three modalities, while three different research participants were assigned to each modality individually. Research participants and researchers completed surveys measuring workload and usability. Remote consenting allows researchers and participants to connect at their preferred location and time, and teleconsenting was found to be the preferred modality by the researchers, primarily due to its ability to exchange visual cues. However, challenges such as training requirements and technology dependence need to be addressed for widespread implementation. Future research should aim to eliminate these barriers and improve remote consenting modalities to facilitate clinical research participation.

Scrybe: A Secure Audit Trail for Clinical Trial Data Fusion.

Clinical trials are a multi-billion dollar industry. One of the biggest challenges facing the clinical trial research community is satisfying Part 11 of Title 21 of the Code of Federal Regulations [7] and ISO 27789 [40]. These controls provide audit requirements that guarantee the reliability of the data contained in the electronic records. Context-aware smart devices and wearable IoT devices have become increasingly common in clinical trials. Electronic Data Capture (EDC) and Clinical Data Management Systems (CDMS) do not currently address the new challenges introduced using these devices. The healthcare digital threat landscape is continually evolving, and the prevalence of sensor fusion and wearable devices compounds the growing attack surface. We propose Scrybe, a permissioned blockchain, to store proof of clinical trial data provenance. We illustrate how Scrybe addresses each control and the limitations of the Ethereum-based blockchains. Finally, we provide a proof-of-concept integration with REDCap to show tamper resistance.

Improving the Quality and Utility of Electronic Health Record Data through Ontologies.

The translational research community, in general, and the Clinical and Translational Science Awards (CTSA) community, in particular, share the vision of repurposing EHRs for research that will improve the quality of clinical practice. Many members of these communities are also aware that electronic health records (EHRs) suffer limitations of data becoming poorly structured, biased, and unusable out of original context. This creates obstacles to the continuity of care, utility, quality improvement, and translational research. Analogous limitations to sharing objective data in other areas of the natural sciences have been successfully overcome by developing and using common ontologies. This White Paper presents the authors' rationale for the use of ontologies with computable semantics for the improvement of clinical data quality and EHR usability formulated for researchers with a stake in clinical and translational science and who are advocates for the use of information technology in medicine but at the same time are concerned by current major shortfalls. This White Paper outlines pitfalls, opportunities, and solutions and recommends increased investment in research and development of ontologies with computable semantics for a new generation of EHRs.

The role of artificial intelligence in hepatology research and practice.

PURPOSE OF REVIEW: The use of artificial intelligence (AI) in examining large data sets has recently gained considerable attention to evaluate disease epidemiology, management approaches, and disease outcomes. The purpose of this review is to summarize the current role of AI in contemporary hepatology practice. RECENT FINDINGS: AI was found to be diagnostically valuable in the evaluation of liver fibrosis, detection of cirrhosis, differentiation between compensated and decompensated cirrhosis, evaluation of portal hypertension, detection and differentiation of particular liver masses, preoperative evaluation of hepatocellular carcinoma as well as response to treatment and estimation of graft survival in patients undergoing liver transplantation. AI additionally holds great promise in examination of structured electronic health records data as well as in examination of clinical text (using various natural language processing approaches). Despite its contributions, AI has several limitations, including the quality of existing data, small cohorts with possible sampling bias and the lack of well validated easily reproducible models. SUMMARY: AI and deep learning models have extensive applicability in assessing liver disease. However, multicenter randomized controlled trials are indispensable to validate their utility.

Contemporary patterns of lipoprotein(a) testing and associated clinical care and outcomes.

Objective: Elevated lipoprotein(a) [Lp(a)] is associated with atherosclerotic cardiovascular disease, yet little is known about Lp(a) testing patterns in real-world practice. The objective of this analysis was to determine how Lp(a) testing is used in clinical practice in comparison with low density lipoprotein cholesterol (LDL-C) testing alone, and to determine whether elevated Lp(a) level is associated with subsequent initiation of lipid-lowering therapy (LLT) and incident cardiovascular (CV) events. Methods: This is an observational cohort study, based on lab tests administered between Jan 1, 2015 and Dec 31, 2019. We used electronic health record (EHR) data from 11 United States health systems participating in the National Patient-Centered Clinical Research Network (PCORnet). We created two cohorts for comparison: 1) the Lp(a) cohort, of adults with an Lp(a) test and 2) the LDL-C cohort, of 4:1 date- and site-matched adults with an LDL-C test, but no Lp(a) test. The primary exposure was the presence of an Lp(a) or LDL-C test result. In the Lp(a) cohort, we used logistic regression to assess the relationship between Lp(a) results in mass units (< 50, 50-100, and > 100mg/dL) and molar units (<125, 125-250, > 250nmol/L) and initiation of LLT within 3 months. We used multivariable adjusted Cox proportional hazards regression to evaluate these Lp(a) levels and time to composite CV hospitalization, including hospitalization for myocardial infarction, revascularization and ischemic stroke. Results: Overall, 20,551 patients had Lp(a) test results and 2,584,773 patients had LDL-C test results (82,204 included in the matched LDL-C cohort). Compared with the LDL-C cohort, the Lp(a) cohort more frequently had prevalent ASCVD (24.3% vs. 8.5%) and multiple prior CV events (8.6% vs. 2.6%). Elevated Lp(a) was associated with greater odds of subsequent LLT initiation. Elevated Lp(a) reported in mass units was also associated with subsequent composite CV hospitalization [aHR (95% CI): Lp(a) 50-100mg/dL 1.25 (1.02-1.53), p<0.03, Lp(a) > 100mg/dL 1.23 (1.08-1.40), p<0.01]. Conclusion: Lp(a) testing is relatively infrequent in health systems across the U.S. As new therapies for Lp(a) emerge, improved patient and provider education is needed to increase awareness of the utility of this risk marker.

Accurate Identification of Patients with Cirrhosis and Its Complications in the Electronic Health Record.

BACKGROUND: Cirrhosis represents a significant health burden; administrative data provide an important tool for research studies. AIMS: We aimed to understand the validity of current ICD-10 codes compared to previously used ICD-9 codes to identify patients with cirrhosis and its complications. METHODS: We identified 1981 patients presenting to MUSC between 2013 and 2019 with a diagnosis of cirrhosis. To validate the sensitivity of ICD codes, we reviewed the medical records of 200 patients for each associated ICD 9 and 10 codes. Sensitivity, specificity, and positive predictive value for each ICD code (individually or when combined) were calculated and univariate binary logistic models, for cirrhosis and its complications, predicted probabilities were used to calculate C-statistics. RESULTS: Single ICD 9 and 10 codes were similarly insensitive for detection of cirrhosis, with sensitivity ranging from 5 to 94%. However, ICD-9 code combinations (when used as either/or) had high sensitivity and specificity for the detection of cirrhosis, with the combination of either 571.5 (or 456.21) or 571.2 codes having a C-statistic of 0.975. Combinations of ICD-10 codes were only slightly less sensitive and specific than ICD-9 codes for detection of cirrhosis (K76.6, or K70.31, plus K74.60 or K74.69, and K70.30 had a C-statistic of 0.927). CONCLUSIONS: ICD-9 and ICD-10 codes when used alone were inaccurate for identifying cirrhosis. ICD-10 and ICD-9 codes had similar performance characteristics. Combinations of ICD codes exhibited the greatest sensitivity and specificity for detection of cirrhosis, and thus should be used to accurately identify cirrhosis.

The effectiveness of a noninterruptive alert to increase prescription of take-home naloxone in emergency departments.

OBJECTIVE: Opioid-related overdose (OD) deaths continue to increase. Take-home naloxone (THN), after treatment for an OD in an emergency department (ED), is a recommended but under-utilized practice. To promote THN prescription, we developed a noninterruptive decision support intervention that combined a detailed OD documentation template with a reminder to use the template that is automatically inserted into a provider's note by decision rules. We studied the impact of the combined intervention on THN prescribing in a longitudinal observational study. METHODS: ED encounters involving an OD were reviewed before and after implementation of the reminder embedded in the physicians' note to use an advanced OD documentation template for changes in: (1) use of the template and (2) prescription of THN. Chi square tests and interrupted time series analyses were used to assess the impact. Usability and satisfaction were measured using the System Usability Scale (SUS) and the Net Promoter Score. RESULTS: In 736 OD cases defined by International Classification of Disease version 10 diagnosis codes (247 prereminder and 489 postreminder), the documentation template was used in 0.0% and 21.3%, respectively (P < .0001). The sensitivity and specificity of the reminder for OD cases were 95.9% and 99.8%, respectively. Use of the documentation template led to twice the rate of prescribing of THN (25.7% vs 50.0%, P < .001). Of 19 providers responding to the survey, 74% of SUS responses were in the good-to-excellent range and 53% of providers were Net Promoters. CONCLUSIONS: A noninterruptive decision support intervention was associated with higher THN prescribing in a pre-post study across a multiinstitution health system.

Red Flag/Blue Flag visualization of a common CNN for text classification.

A shallow convolutional neural network (CNN), TextCNN, has become nearly ubiquitous for classification among clinical and medical text. This research presents a novel eXplainable-AI (X-AI) software, Red Flag/Blue Flag (RFBF), designed for binary classification with TextCNN. RFBF visualizes each convolutional filter's discriminative capability. This is a more informative approach than direct assessment of logit contribution, features that overfit to train set nuances on smaller datasets may indiscriminately activate large logits on validation samples from both classes. RFBF enables model diagnosis, term feature verification, and overfit prevention. We present 3 use cases of (1) filter consistency assessment; (2) predictive performance improvement; and (3) estimation of information leakage between train and holdout sets. The use cases derive from experiments on TextCNN for binary prediction of surgical misadventure outcomes from physician-authored operative notes. Due to TextCNN's prevalence, this X-AI can benefit clinical text research, and hence improve patient outcomes.

An AI Approach for Identifying Patients With Cirrhosis.

GOAL: The goal of this study was to evaluate an artificial intelligence approach, namely deep learning, on clinical text in electronic health records (EHRs) to identify patients with cirrhosis. BACKGROUND AND AIMS: Accurate identification of cirrhosis in EHR is important for epidemiological, health services, and outcomes research. Currently, such efforts depend on International Classification of Diseases (ICD) codes, with limited success. MATERIALS AND METHODS: We trained several machine learning models using discharge summaries from patients with known cirrhosis from a patient registry and random controls without cirrhosis or its complications based on ICD codes. Models were validated on patients for whom discharge summaries were manually reviewed and used as the gold standard test set. We tested Naive Bayes and Random Forest as baseline models and a deep learning model using word embedding and a convolutional neural network (CNN). RESULTS: The training set included 446 cirrhosis patients and 689 controls, while the gold standard test set included 139 cirrhosis patients and 152 controls. Among the machine learning models, the CNN achieved the highest area under the receiver operating characteristic curve (0.993), with a precision of 0.965 and recall of 0.978, compared with 0.879 and 0.981 for the Naive Bayes and Random Forest, respectively (precision 0.787 and 0.958, and recalls 0.878 and 0.827). The precision by ICD codes for cirrhosis was 0.883 and recall was 0.978. CONCLUSIONS: A CNN model trained on discharge summaries identified cirrhosis patients with high precision and recall. This approach for phenotyping cirrhosis in the EHR may provide a more accurate assessment of disease burden in a variety of studies.

Enhanced phenotypes for identifying opioid overdose in emergency department visit electronic health record data.

Background: Accurate identification of opioid overdose (OOD) cases in electronic healthcare record (EHR) data is an important element in surveillance, empirical research, and clinical intervention. We sought to improve existing OOD electronic phenotypes by incorporating new data types beyond diagnostic codes and by applying several statistical and machine learning methods. Materials and Methods: We developed an EHR dataset of emergency department visits involving OOD cases or patients considered at risk for an OOD and ascertained true OOD status through manual chart reviews. We developed and validated prediction models using Random Forest, Extreme Gradient Boost, and Elastic Net models that incorporated 717 features involving primary and second diagnoses, chief complaints, medications prescribed, vital signs, laboratory results, and procedural codes. We also developed models limited to single data types. Results: A total of 1718 records involving 1485 patients were manually reviewed; 541 (36.4%) patients had one or more OOD. Prediction performance was similar for all models; sensitivity varied from 94% to 97%; and area under the receiver operating characteristic curve (AUC) was 98% for all methods. The primary diagnosis and chief complaint were the most important contributors to AUC performance; primary diagnoses and medication class contributed most to sensitivity; chief complaint, primary diagnosis, and vital signs were most important for specificity. Models limited to decision support data types available in real time demonstrated robust prediction performance. Conclusions: Substantial prediction performance improvements were demonstrated for identifying OODs in EHR data. Our e-phenotypes could be applied in surveillance, retrospective empirical applications, or clinical decision support systems.

Enhancing research data infrastructure to address the opioid epidemic: the Opioid Overdose Network (O2-Net).

Opioid Overdose Network is an effort to generalize and adapt an existing research data network, the Accrual to Clinical Trials (ACT) Network, to support design of trials for survivors of opioid overdoses presenting to emergency departments (ED). Four institutions (Medical University of South Carolina [MUSC], Dartmouth Medical School [DMS], University of Kentucky [UK], and University of California San Diego [UCSD]) worked to adapt the ACT network. The approach that was taken to enhance the ACT network focused on 4 activities: cloning and extending the ACT infrastructure, developing an e-phenotype and corresponding registry, developing portable natural language processing tools to enhance data capture, and developing automated documentation templates to enhance extended data capture. Overall, initial results suggest that tailoring of existing multipurpose federated research networks to specific tasks is feasible; however, substantial efforts are required for coordination of the subnetwork and development of new tools for extension of available data. The initial output of the project was a new approach to decision support for the prescription of naloxone for home use in the ED, which is under further study within the network.

Natural language processing enabling COVID-19 predictive analytics to support data-driven patient advising and pooled testing.

OBJECTIVE: The COVID-19 (coronavirus disease 2019) pandemic response at the Medical University of South Carolina included virtual care visits for patients with suspected severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) infection. The telehealth system used for these visits only exports a text note to integrate with the electronic health record, but structured and coded information about COVID-19 (eg, exposure, risk factors, symptoms) was needed to support clinical care and early research as well as predictive analytics for data-driven patient advising and pooled testing. MATERIALS AND METHODS: To capture COVID-19 information from multiple sources, a new data mart and a new natural language processing (NLP) application prototype were developed. The NLP application combined reused components with dictionaries and rules crafted by domain experts. It was deployed as a Web service for hourly processing of new data from patients assessed or treated for COVID-19. The extracted information was then used to develop algorithms predicting SARS-CoV-2 diagnostic test results based on symptoms and exposure information. RESULTS: The dedicated data mart and NLP application were developed and deployed in a mere 10-day sprint in March 2020. The NLP application was evaluated with good accuracy (85.8% recall and 81.5% precision). The SARS-CoV-2 testing predictive analytics algorithms were configured to provide patients with data-driven COVID-19 testing advices with a sensitivity of 81% to 92% and to enable pooled testing with a negative predictive value of 90% to 91%, reducing the required tests to about 63%. CONCLUSIONS: SARS-CoV-2 testing predictive analytics and NLP successfully enabled data-driven patient advising and pooled testing.

Racial Differences in Patient Portal Activation and Research Enrollment Among Patients With Prostate Cancer.

PURPOSE: The purpose of this study was to examine racial differences in patient portal activation and research participation among patients with prostate cancer. MATERIALS AND METHODS: Participants were African American and White patients with prostate cancer who were treated with radical prostatectomy (n = 218). Patient portal activation was determined using electronic health records, and research participation was measured based on completion of a social determinants survey. RESULTS: Thirty-one percent of patients completed the social determinants survey and enrolled in the study and 66% activated a patient portal. The likelihood of enrolling in the study was reduced with greater levels of social deprivation (odds ratio [OR], 0.70; 95% CI, 0.50 to 0.98; P = .04). Social deprivation also had a signification independent association with patient portal activation along with racial background. African American patients (OR, 0.48; 95% CI, 0.23 to 0.91; P = .02) and those with greater social deprivation (OR, 0.58; 95% CI, 0.42 to 0.82; P = .002) had a lower likelihood of activating a patient portal compared with White patients and those with lower social deprivation. CONCLUSION: Although the majority of patients with prostate cancer activated their patient portal, rates of patient portal activation were lower among African American patients and those who lived in areas with greater social deprivation. Greater efforts are needed to promote patient portal activation among African American patients with prostate cancer and address access to health information technology among those who live in socially disadvantaged geographic areas.

Feasibility and Impact of the Combined Application of Coronary CT Angiography With the HEART Pathway in Patients With Suspected Acute Coronary Syndrome.

BACKGROUND: This study examines the feasibility and utility of integrating coronary computed tomography angiography and the HEART Pathway into a novel accelerated diagnostic protocol-called HEART-CT-and assesses its impact as an optional interactive decision support tool (smart form) in the electronic health record. METHODS: This was a retrospective observational study performed in 2 adult emergency departments (ED) among patients evaluated for suspected acute coronary syndrome. Primary outcomes included the rate of discharge from the ED following HEART-CT Smart Form use, 30-day major adverse cardiac events (MACE), and ED length of stay (LOS). Hypothesis-generating outcomes included the rate of Smart Form use by ED providers and whether adhering to the HEART-CT recommendations was associated with improved outcomes. RESULTS: The study included 672 subjects, 78.1% of whom were discharged from the ED. HEART-CT identified 76.7% of patients with increased risk HEAR scores as safe for discharge. No patients identified as low risk by HEART-CT had MACE within 30 days. Total mean ED LOS was 4.6 hours. ED providers used the HEART-CT smart form in 19.7% of eligible patients. ED providers who followed the HEART-CT recommendations had 3.41 times higher odds of ED discharging patients with increased risk HEAR scores than nonadherent providers (95% CI, 2.20-5.27). CONCLUSIONS: HEART-CT reclassified a large proportion of patients as safe for discharge, maintained a high sensitivity for detecting 30-day MACE, and had an acceptable ED LOS. Future studies should test the extent to which more automated clinical decision support improves provider adoption and clinical outcomes of HEART-CT.

Research Integrated Network of Systems (RINS): a virtual data warehouse for the acceleration of translational research.

OBJECTIVE: Integrated, real-time data are crucial to evaluate translational efforts to accelerate innovation into care. Too often, however, needed data are fragmented in disparate systems. The South Carolina Clinical & Translational Research Institute at the Medical University of South Carolina (MUSC) developed and implemented a universal study identifier-the Research Master Identifier (RMID)-for tracking research studies across disparate systems and a data warehouse-inspired model-the Research Integrated Network of Systems (RINS)-for integrating data from those systems. MATERIALS AND METHODS: In 2017, MUSC began requiring the use of RMIDs in informatics systems that support human subject studies. We developed a web-based tool to create RMIDs and application programming interfaces to synchronize research records and visualize linkages to protocols across systems. Selected data from these disparate systems were extracted and merged nightly into an enterprise data mart, and performance dashboards were created to monitor key translational processes. RESULTS: Within 4 years, 5513 RMIDs were created. Among these were 726 (13%) bridged systems needed to evaluate research study performance, and 982 (18%) linked to the electronic health records, enabling patient-level reporting. DISCUSSION: Barriers posed by data fragmentation to assessment of program impact have largely been eliminated at MUSC through the requirement for an RMID, its distribution via RINS to disparate systems, and mapping of system-level data to a single integrated data mart. CONCLUSION: By applying data warehousing principles to federate data at the "study" level, the RINS project reduced data fragmentation and promoted research systems integration.

Cloud-Based Federated Learning Implementation Across Medical Centers.

PURPOSE: Building well-performing machine learning (ML) models in health care has always been exigent because of the data-sharing concerns, yet ML approaches often require larger training samples than is afforded by one institution. This paper explores several federated learning implementations by applying them in both a simulated environment and an actual implementation using electronic health record data from two academic medical centers on a Microsoft Azure Cloud Databricks platform. MATERIALS AND METHODS: Using two separate cloud tenants, ML models were created, trained, and exchanged from one institution to another via a GitHub repository. Federated learning processes were applied to both artificial neural networks (ANNs) and logistic regression (LR) models on the horizontal data sets that are varying in count and availability. Incremental and cyclic federated learning models have been tested in simulation and real environments. RESULTS: The cyclically trained ANN showed a 3% increase in performance, a significant improvement across most attempts (P < .05). Single weight neural network models showed improvement in some cases. However, LR models did not show much improvement after federated learning processes. The specific process that improved the performance differed based on the ML model and how federated learning was implemented. Moreover, we have confirmed that the order of the institutions during the training did influence the overall performance increase. CONCLUSION: Unlike previous studies, our work has shown the implementation and effectiveness of federated learning processes beyond simulation. Additionally, we have identified different federated learning models that have achieved statistically significant performances. More work is needed to achieve effective federated learning processes in biomedicine, while preserving the security and privacy of the data.

Each patient is a research biorepository: informatics-enabled research on surplus clinical specimens via the living BioBank.

The ability to analyze human specimens is the pillar of modern-day translational research. To enhance the research availability of relevant clinical specimens, we developed the Living BioBank (LBB) solution, which allows for just-in-time capture and delivery of phenotyped surplus laboratory medicine specimens. The LBB is a system-of-systems integrating research feasibility databases in i2b2, a real-time clinical data warehouse, and an informatics system for institutional research services management (SPARC). LBB delivers deidentified clinical data and laboratory specimens. We further present an extension to our solution, the Living µBiome Bank, that allows the user to request and receive phenotyped specimen microbiome data. We discuss the details of the implementation of the LBB system and the necessary regulatory oversight for this solution. The conducted institutional focus group of translational investigators indicates an overall positive sentiment towards potential scientific results generated with the use of LBB. Reference implementation of LBB is available at https://LivingBioBank.musc.edu.

International electronic health record-derived COVID-19 clinical course profiles: the 4CE consortium.

We leveraged the largely untapped resource of electronic health record data to address critical clinical and epidemiological questions about Coronavirus Disease 2019 (COVID-19). To do this, we formed an international consortium (4CE) of 96 hospitals across five countries (www.covidclinical.net). Contributors utilized the Informatics for Integrating Biology and the Bedside (i2b2) or Observational Medical Outcomes Partnership (OMOP) platforms to map to a common data model. The group focused on temporal changes in key laboratory test values. Harmonized data were analyzed locally and converted to a shared aggregate form for rapid analysis and visualization of regional differences and global commonalities. Data covered 27,584 COVID-19 cases with 187,802 laboratory tests. Case counts and laboratory trajectories were concordant with existing literature. Laboratory tests at the time of diagnosis showed hospital-level differences equivalent to country-level variation across the consortium partners. Despite the limitations of decentralized data generation, we established a framework to capture the trajectory of COVID-19 disease in patients and their response to interventions.

Identifying and Predicting Intentional Self-Harm in Electronic Health Record Clinical Notes: Deep Learning Approach.

BACKGROUND: Suicide is an important public health concern in the United States and around the world. There has been significant work examining machine learning approaches to identify and predict intentional self-harm and suicide using existing data sets. With recent advances in computing, deep learning applications in health care are gaining momentum. OBJECTIVE: This study aimed to leverage the information in clinical notes using deep neural networks (DNNs) to (1) improve the identification of patients treated for intentional self-harm and (2) predict future self-harm events. METHODS: We extracted clinical text notes from electronic health records (EHRs) of 835 patients with International Classification of Diseases (ICD) codes for intentional self-harm and 1670 matched controls who never had any intentional self-harm ICD codes. The data were divided into training and holdout test sets. We tested a number of algorithms on clinical notes associated with the intentional self-harm codes using the training set, including several traditional bag-of-words-based models and 2 DNN models: a convolutional neural network (CNN) and a long short-term memory model. We also evaluated the predictive performance of the DNNs on a subset of patients who had clinical notes 1 to 6 months before the first intentional self-harm event. Finally, we evaluated the impact of a pretrained model using Word2vec (W2V) on performance. RESULTS: The area under the receiver operating characteristic curve (AUC) for the CNN on the phenotyping task, that is, the detection of intentional self-harm in clinical notes concurrent with the events was 0.999, with an F1 score of 0.985. In the predictive task, the CNN achieved the highest performance with an AUC of 0.882 and an F1 score of 0.769. Although pretraining with W2V shortened the DNN training time, it did not improve performance. CONCLUSIONS: The strong performance on the first task, namely, phenotyping based on clinical notes, suggests that such models could be used effectively for surveillance of intentional self-harm in clinical text in an EHR. The modest performance on the predictive task notwithstanding, the results using DNN models on clinical text alone are competitive with other reports in the literature using risk factors from structured EHR data.