Identifying social determinants of health from clinical narratives: A study of performance, documentation ratio, and potential bias.

OBJECTIVE: To develop a natural language processing (NLP) package to extract social determinants of health (SDoH) from clinical narratives, examine the bias among race and gender groups, test the generalizability of extracting SDoH for different disease groups, and examine population-level extraction ratio. METHODS: We developed SDoH corpora using clinical notes identified at the University of Florida (UF) Health. We systematically compared 7 transformer-based large language models (LLMs) and developed an open-source package - SODA (i.e., SOcial DeterminAnts) to facilitate SDoH extraction from clinical narratives. We examined the performance and potential bias of SODA for different race and gender groups, tested the generalizability of SODA using two disease domains including cancer and opioid use, and explored strategies for improvement. We applied SODA to extract 19 categories of SDoH from the breast (n = 7,971), lung (n = 11,804), and colorectal cancer (n = 6,240) cohorts to assess patient-level extraction ratio and examine the differences among race and gender groups. RESULTS: We developed an SDoH corpus using 629 clinical notes of cancer patients with annotations of 13,193 SDoH concepts/attributes from 19 categories of SDoH, and another cross-disease validation corpus using 200 notes from opioid use patients with 4,342 SDoH concepts/attributes. We compared 7 transformer models and the GatorTron model achieved the best mean average strict/lenient F1 scores of 0.9122 and 0.9367 for SDoH concept extraction and 0.9584 and 0.9593 for linking attributes to SDoH concepts. There is a small performance gap (∼4%) between Males and Females, but a large performance gap (>16 %) among race groups. The performance dropped when we applied the cancer SDoH model to the opioid cohort; fine-tuning using a smaller opioid SDoH corpus improved the performance. The extraction ratio varied in the three cancer cohorts, in which 10 SDoH could be extracted from over 70 % of cancer patients, but 9 SDoH could be extracted from less than 70 % of cancer patients. Individuals from the White and Black groups have a higher extraction ratio than other minority race groups. CONCLUSIONS: Our SODA package achieved good performance in extracting 19 categories of SDoH from clinical narratives. The SODA package with pre-trained transformer models is available at https://github.com/uf-hobi-informatics-lab/SODA_Docker.

Detection of Personal and Family History of Suicidal Thoughts and Behaviors using Deep Learning and Natural Language Processing: A Multi-Site Study.

Objective: Personal and family history of suicidal thoughts and behaviors (PSH and FSH, respectively) are significant risk factors associated with future suicide events. These are often captured in narrative clinical notes in electronic health records (EHRs). Collaboratively, Weill Cornell Medicine (WCM), Northwestern Medicine (NM), and the University of Florida (UF) developed and validated deep learning (DL)-based natural language processing (NLP) tools to detect PSH and FSH from such notes. The tool's performance was further benchmarked against a method relying exclusively on ICD-9/10 diagnosis codes. Materials and Methods: We developed DL-based NLP tools utilizing pre-trained transformer models Bio_ClinicalBERT and GatorTron, and compared them with expert-informed, rule-based methods. The tools were initially developed and validated using manually annotated clinical notes at WCM. Their portability and performance were further evaluated using clinical notes at NM and UF. Results: The DL tools outperformed the rule-based NLP tool in identifying PSH and FHS. For detecting PSH, the rule-based system obtained an F1-score of 0.75 ± 0.07, while the Bio_ClinicalBERT and GatorTron DL tools scored 0.83 ± 0.09 and 0.84 ± 0.07, respectively. For detecting FSH, the rule-based NLP tool's F1-score was 0.69 ± 0.11, compared to 0.89 ± 0.10 for Bio_ClinicalBERT and 0.92 ± 0.07 for GatorTron. For the gold standard corpora across the three sites, only 2.2% (WCM), 9.3% (NM), and 7.8% (UF) of patients reported to have an ICD-9/10 diagnosis code for suicidal thoughts and behaviors prior to the clinical notes report date. The best performing GatorTron DL tool identified 93.0% (WCM), 80.4% (NM), and 89.0% (UF) of patients with documented PSH, and 85.0%(WCM), 89.5%(NM), and 100%(UF) of patients with documented FSH in their notes. Discussion: While PSH and FSH are significant risk factors for future suicide events, little effort has been made previously to identify individuals with these history. To address this, we developed a transformer based DL method and compared with conventional rule-based NLP approach. The varying effectiveness of the rule-based tools across sites suggests a need for improvement in its dictionary-based approach. In contrast, the performances of the DL tools were higher and comparable across sites. Furthermore, DL tools were fine-tuned using only small number of annotated notes at each site, underscores its greater adaptability to local documentation practices and lexical variations. Conclusion: Variations in local documentation practices across health care systems pose challenges to rule-based NLP tools. In contrast, the developed DL tools can effectively extract PSH and FSH information from unstructured clinical notes. These tools will provide clinicians with crucial information for assessing and treating patients at elevated risk for suicide who are rarely been diagnosed.

Using Machine Learning to Predict Antidepressant Treatment Outcome From Electronic Health Records.

Objective: To evaluate if a machine learning approach can accurately predict antidepressant treatment outcome using electronic health records (EHRs) from patients with depression. Method: This study examined 808 patients with depression at a New York City-based outpatient mental health clinic between June 13, 2016 and June 22, 2020. Antidepressant treatment outcome was defined based on trend in depression symptom severity over time and was categorized as either "Recovering" or "Worsening" (i.e., non-Recovering), measured by the slope of individual-level Patient Health Questionnaire-9 (PHQ-9) score trajectory spanning 6 months following treatment initiation. A patient was designated as "Recovering" if the slope is less than 0 and as "Worsening" if the slope was no less than 0. Multiple machine learning (ML) models including L2 norm regularized Logistic Regression, Naive Bayes, Random Forest, and Gradient Boosting Decision Tree (GBDT) were used to predict treatment outcome based on additional data from EHRs, including demographics and diagnoses. Shapley Additive Explanations were applied to identify the most important predictors. Results: The GBDT achieved the best results of predicting "Recovering" (AUC: 0.7654 ± 0.0227; precision: 0.6002 ± 0.0215; recall: 0.5131 ± 0.0336). When excluding patients with low PHQ-9 scores (<10) at baseline, the results of predicting "Recovering" (AUC: 0.7254 ± 0.0218; precision: 0.5392 ± 0.0437; recall: 0.4431 ± 0.0513) were obtained. Prior diagnosis of anxiety, psychotherapy, recurrent depression, and baseline depression symptom severity were strong predictors. Conclusions: The results demonstrate the potential utility of using ML in longitudinal EHRs to predict antidepressant treatment outcome. Our predictive tool holds the promise to accelerate personalized medical management in patients with psychiatric illnesses.

AD-BERT: Using pre-trained language model to predict the progression from mild cognitive impairment to Alzheimer's disease.

OBJECTIVE: We develop a deep learning framework based on the pre-trained Bidirectional Encoder Representations from Transformers (BERT) model using unstructured clinical notes from electronic health records (EHRs) to predict the risk of disease progression from Mild Cognitive Impairment (MCI) to Alzheimer's Disease (AD). METHODS: We identified 3657 patients diagnosed with MCI together with their progress notes from Northwestern Medicine Enterprise Data Warehouse (NMEDW) between 2000 and 2020. The progress notes no later than the first MCI diagnosis were used for the prediction. We first preprocessed the notes by deidentification, cleaning and splitting into sections, and then pre-trained a BERT model for AD (named AD-BERT) based on the publicly available Bio+Clinical BERT on the preprocessed notes. All sections of a patient were embedded into a vector representation by AD-BERT and then combined by global MaxPooling and a fully connected network to compute the probability of MCI-to-AD progression. For validation, we conducted a similar set of experiments on 2563 MCI patients identified at Weill Cornell Medicine (WCM) during the same timeframe. RESULTS: Compared with the 7 baseline models, the AD-BERT model achieved the best performance on both datasets, with Area Under receiver operating characteristic Curve (AUC) of 0.849 and F1 score of 0.440 on NMEDW dataset, and AUC of 0.883 and F1 score of 0.680 on WCM dataset. CONCLUSION: The use of EHRs for AD-related research is promising, and AD-BERT shows superior predictive performance in modeling MCI-to-AD progression prediction. Our study demonstrates the utility of pre-trained language models and clinical notes in predicting MCI-to-AD progression, which could have important implications for improving early detection and intervention for AD.

Comparing the effects of four common drug classes on the progression of mild cognitive impairment to dementia using electronic health records.

The objective of this study was to investigate the potential association between the use of four frequently prescribed drug classes, namely antihypertensive drugs, statins, selective serotonin reuptake inhibitors, and proton-pump inhibitors, and the likelihood of disease progression from mild cognitive impairment (MCI) to dementia using electronic health records (EHRs). We conducted a retrospective cohort study using observational EHRs from a cohort of approximately 2 million patients seen at a large, multi-specialty urban academic medical center in New York City, USA between 2008 and 2020 to automatically emulate the randomized controlled trials. For each drug class, two exposure groups were identified based on the prescription orders documented in the EHRs following their MCI diagnosis. During follow-up, we measured drug efficacy based on the incidence of dementia and estimated the average treatment effect (ATE) of various drugs. To ensure the robustness of our findings, we confirmed the ATE estimates via bootstrapping and presented associated 95% confidence intervals (CIs). Our analysis identified 14,269 MCI patients, among whom 2501 (17.5%) progressed to dementia. Using average treatment estimation and bootstrapping confirmation, we observed that drugs including rosuvastatin (ATE = - 0.0140 [- 0.0191, - 0.0088], p value < 0.001), citalopram (ATE = - 0.1128 [- 0.125, - 0.1005], p value < 0.001), escitalopram (ATE = - 0.0560 [- 0.0615, - 0.0506], p value < 0.001), and omeprazole (ATE = - 0.0201 [- 0.0299, - 0.0103], p value < 0.001) have a statistically significant association in slowing the progression from MCI to dementia. The findings from this study support the commonly prescribed drugs in altering the progression from MCI to dementia and warrant further investigation.

Prediction of left ventricular ejection fraction changes in heart failure patients using machine learning and electronic health records: a multi-site study.

Left ventricular ejection fraction (EF) is a key measure in the diagnosis and treatment of heart failure (HF) and many patients experience changes in EF overtime. Large-scale analysis of longitudinal changes in EF using electronic health records (EHRs) is limited. In a multi-site retrospective study using EHR data from three academic medical centers, we investigated longitudinal changes in EF measurements in patients diagnosed with HF. We observed significant variations in baseline characteristics and longitudinal EF change behavior of the HF cohorts from a previous study that is based on HF registry data. Data gathered from this longitudinal study were used to develop multiple machine learning models to predict changes in ejection fraction measurements in HF patients. Across all three sites, we observed higher performance in predicting EF increase over a 1-year duration, with similarly higher performance predicting an EF increase of 30% from baseline compared to lower percentage increases. In predicting EF decrease we found moderate to high performance with low confidence for various models. Among various machine learning models, XGBoost was the best performing model for predicting EF changes. Across the three sites, the XGBoost model had an F1-score of 87.2, 89.9, and 88.6 and AUC of 0.83, 0.87, and 0.90 in predicting a 30% increase in EF, and had an F1-score of 95.0, 90.6, 90.1 and AUC of 0.54, 0.56, 0.68 in predicting a 30% decrease in EF. Among features that contribute to predicting EF changes, baseline ejection fraction measurement, age, gender, and heart diseases were found to be statistically significant.

Comorbidity and healthcare utilization in patients with treatment resistant depression: A large-scale retrospective cohort analysis using electronic health records.

BACKGROUND: Medical comorbidity and healthcare utilization in patients with treatment resistant depression (TRD) is usually reported in convenience samples, making estimates unreliable. There is only limited large-scale clinical research on comorbidities and healthcare utilization in TRD patients. METHODS: Electronic Health Record data from over 3.3 million patients from the INSIGHT Clinical Research Network in New York City was used to define TRD as initiation of a third antidepressant regimen in a 12-month period among patients diagnosed with major depressive disorder (MDD). Age and sex matched TRD and non-TRD MDD patients were compared for anxiety disorder, 27 comorbid medical conditions, and healthcare utilization. RESULTS: Out of 30,218 individuals diagnosed with MDD, 15.2 % of patients met the criteria for TRD (n = 4605). Compared to MDD patients without TRD, the TRD patients had higher rates of anxiety disorder and physical comorbidities. They also had higher odds of ischemic heart disease (OR = 1.38), stroke/transient ischemic attack (OR = 1.57), chronic kidney diseases (OR = 1.53), arthritis (OR = 1.52), hip/pelvic fractures (OR = 2.14), and cancers (OR = 1.41). As compared to non-TRD MDD, TRD patients had higher rates of emergency room visits, and inpatient stays. In relation to patients without MDD, both TRD and non-TRD MDD patients had significantly higher levels of anxiety disorder and physical comorbidities. LIMITATIONS: The INSIGHT-CRN data lack information on depression severity and medication adherence. CONCLUSIONS: TRD patients compared to non-TRD MDD patients have a substantially higher prevalence of various psychiatric and medical comorbidities and higher health care utilization. These findings highlight the challenges of developing interventions and care coordination strategies to meet the complex clinical needs of TRD patients.

Social connectedness as a determinant of mental health: A scoping review.

Public health and epidemiologic research have established that social connectedness promotes overall health. Yet there have been no recent reviews of findings from research examining social connectedness as a determinant of mental health. The goal of this review was to evaluate recent longitudinal research probing the effects of social connectedness on depression and anxiety symptoms and diagnoses in the general population. A scoping review was performed of PubMed and PsychInfo databases from January 2015 to December 2021 following PRISMA-ScR guidelines using a defined search strategy. The search yielded 66 unique studies. In research with other than pregnant women, 83% (19 of 23) studies reported that social support benefited symptoms of depression with the remaining 17% (5 of 23) reporting minimal or no evidence that lower levels of social support predict depression at follow-up. In research with pregnant women, 83% (24 of 29 studies) found that low social support increased postpartum depressive symptoms. Among 8 of 9 studies that focused on loneliness, feeling lonely at baseline was related to adverse outcomes at follow-up including higher risks of major depressive disorder, depressive symptom severity, generalized anxiety disorder, and lower levels of physical activity. In 5 of 8 reports, smaller social network size predicted depressive symptoms or disorder at follow-up. In summary, most recent relevant longitudinal studies have demonstrated that social connectedness protects adults in the general population from depressive symptoms and disorders. The results, which were largely consistent across settings, exposure measures, and populations, support efforts to improve clinical detection of high-risk patients, including adults with low social support and elevated loneliness.

Predictive Modeling for Suicide-Related Outcomes and Risk Factors among Patients with Pain Conditions: A Systematic Review.

Suicide is a leading cause of death in the US. Patients with pain conditions have higher suicidal risks. In a systematic review searching observational studies from multiple sources (e.g., MEDLINE) from 1 January 2000-12 September 2020, we evaluated existing suicide prediction models' (SPMs) performance and identified risk factors and their derived data sources among patients with pain conditions. The suicide-related outcomes included suicidal ideation, suicide attempts, suicide deaths, and suicide behaviors. Among the 87 studies included (with 8 SPM studies), 107 suicide risk factors (grouped into 27 categories) were identified. The most frequently occurring risk factor category was depression and their severity (33%). Approximately 20% of the risk factor categories would require identification from data sources beyond structured data (e.g., clinical notes). For 8 SPM studies (only 2 performing validation), the reported prediction metrics/performance varied: C-statistics (n = 3 studies) ranged 0.67-0.84, overall accuracy(n = 5): 0.78-0.96, sensitivity(n = 2): 0.65-0.91, and positive predictive values(n = 3): 0.01-0.43. Using the modified Quality in Prognosis Studies tool to assess the risk of biases, four SPM studies had moderate-to-high risk of biases. This systematic review identified a comprehensive list of risk factors that may improve predicting suicidal risks for patients with pain conditions. Future studies need to examine reasons for performance variations and SPM's clinical utility.

Design and validation of a FHIR-based EHR-driven phenotyping toolbox.

OBJECTIVES: To develop and validate a standards-based phenotyping tool to author electronic health record (EHR)-based phenotype definitions and demonstrate execution of the definitions against heterogeneous clinical research data platforms. MATERIALS AND METHODS: We developed an open-source, standards-compliant phenotyping tool known as the PhEMA Workbench that enables a phenotype representation using the Fast Healthcare Interoperability Resources (FHIR) and Clinical Quality Language (CQL) standards. We then demonstrated how this tool can be used to conduct EHR-based phenotyping, including phenotype authoring, execution, and validation. We validated the performance of the tool by executing a thrombotic event phenotype definition at 3 sites, Mayo Clinic (MC), Northwestern Medicine (NM), and Weill Cornell Medicine (WCM), and used manual review to determine precision and recall. RESULTS: An initial version of the PhEMA Workbench has been released, which supports phenotype authoring, execution, and publishing to a shared phenotype definition repository. The resulting thrombotic event phenotype definition consisted of 11 CQL statements, and 24 value sets containing a total of 834 codes. Technical validation showed satisfactory performance (both NM and MC had 100% precision and recall and WCM had a precision of 95% and a recall of 84%). CONCLUSIONS: We demonstrate that the PhEMA Workbench can facilitate EHR-driven phenotype definition, execution, and phenotype sharing in heterogeneous clinical research data environments. A phenotype definition that integrates with existing standards-compliant systems, and the use of a formal representation facilitates automation and can decrease potential for human error.

Extracting social determinants of health from electronic health records using natural language processing: a systematic review.

OBJECTIVE: Social determinants of health (SDoH) are nonclinical dispositions that impact patient health risks and clinical outcomes. Leveraging SDoH in clinical decision-making can potentially improve diagnosis, treatment planning, and patient outcomes. Despite increased interest in capturing SDoH in electronic health records (EHRs), such information is typically locked in unstructured clinical notes. Natural language processing (NLP) is the key technology to extract SDoH information from clinical text and expand its utility in patient care and research. This article presents a systematic review of the state-of-the-art NLP approaches and tools that focus on identifying and extracting SDoH data from unstructured clinical text in EHRs. MATERIALS AND METHODS: A broad literature search was conducted in February 2021 using 3 scholarly databases (ACL Anthology, PubMed, and Scopus) following Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines. A total of 6402 publications were initially identified, and after applying the study inclusion criteria, 82 publications were selected for the final review. RESULTS: Smoking status (n = 27), substance use (n = 21), homelessness (n = 20), and alcohol use (n = 15) are the most frequently studied SDoH categories. Homelessness (n = 7) and other less-studied SDoH (eg, education, financial problems, social isolation and support, family problems) are mostly identified using rule-based approaches. In contrast, machine learning approaches are popular for identifying smoking status (n = 13), substance use (n = 9), and alcohol use (n = 9). CONCLUSION: NLP offers significant potential to extract SDoH data from narrative clinical notes, which in turn can aid in the development of screening tools, risk prediction models, and clinical decision support systems.

Automated Extraction of Tumor Staging and Diagnosis Information From Surgical Pathology Reports.

PURPOSE: Typically stored as unstructured notes, surgical pathology reports contain data elements valuable to cancer research that require labor-intensive manual extraction. Although studies have described natural language processing (NLP) of surgical pathology reports to automate information extraction, efforts have focused on specific cancer subtypes rather than across multiple oncologic domains. To address this gap, we developed and evaluated an NLP method to extract tumor staging and diagnosis information across multiple cancer subtypes. METHODS: The NLP pipeline was implemented on an open-source framework called Leo. We used a total of 555,681 surgical pathology reports of 329,076 patients to develop the pipeline and evaluated our approach on subsets of reports from patients with breast, prostate, colorectal, and randomly selected cancer subtypes. RESULTS: Averaged across all four cancer subtypes, the NLP pipeline achieved an accuracy of 1.00 for International Classification of Diseases, Tenth Revision codes, 0.89 for T staging, 0.90 for N staging, and 0.97 for M staging. It achieved an F1 score of 1.00 for International Classification of Diseases, Tenth Revision codes, 0.88 for T staging, 0.90 for N staging, and 0.24 for M staging. CONCLUSION: The NLP pipeline was developed to extract tumor staging and diagnosis information across multiple cancer subtypes to support the research enterprise in our institution. Although it was not possible to demonstrate generalizability of our NLP pipeline to other institutions, other institutions may find value in adopting a similar NLP approach-and reusing code available at GitHub-to support the oncology research enterprise with elements extracted from surgical pathology reports.

CQL4NLP: Development and Integration of FHIR NLP Extensions in Clinical Quality Language for EHR-driven Phenotyping.

Lack of standardized representation of natural language processing (NLP) components in phenotyping algorithms hinders portability of the phenotyping algorithms and their execution in a high-throughput and reproducible manner. The objective of the study is to develop and evaluate a standard-driven approach - CQL4NLP - that integrates a collection of NLP extensions represented in the HL7 Fast Healthcare Interoperability Resources (FHIR) standard into the clinical quality language (CQL). A minimal NLP data model with 11 NLP-specific data elements was created, including six FHIR NLP extensions. All 11 data elements were identified from their usage in real-world phenotyping algorithms. An NLP ruleset generation mechanism was integrated into the NLP2FHIR pipeline and the NLP rulesets enabled comparable performance for a case study with the identification of obesity comorbidities. The NLP ruleset generation mechanism created a reproducible process for defining the NLP components of a phenotyping algorithm and its execution.

Using weak supervision and deep learning to classify clinical notes for identification of current suicidal ideation.

Mental health concerns, such as suicidal thoughts, are frequently documented by providers in clinical notes, as opposed to structured coded data. In this study, we evaluated weakly supervised methods for detecting "current" suicidal ideation from unstructured clinical notes in electronic health record (EHR) systems. Weakly supervised machine learning methods leverage imperfect labels for training, alleviating the burden of creating a large manually annotated dataset. After identifying a cohort of 600 patients at risk for suicidal ideation, we used a rule-based natural language processing approach (NLP) approach to label the training and validation notes (n = 17,978). Using this large corpus of clinical notes, we trained several statistical machine learning models-logistic classifier, support vector machines (SVM), Naive Bayes classifier-and one deep learning model, namely a text classification convolutional neural network (CNN), to be evaluated on a manually-reviewed test set (n = 837). The CNN model outperformed all other methods, achieving an overall accuracy of 94% and a F1-score of 0.82 on documents with "current" suicidal ideation. This algorithm correctly identified an additional 42 encounters and 9 patients indicative of suicidal ideation but missing a structured diagnosis code. When applied to a random subset of 5,000 clinical notes, the algorithm classified 0.46% (n = 23) for "current" suicidal ideation, of which 87% were truly indicative via manual review. Implementation of this approach for large-scale document screening may play an important role in point-of-care clinical information systems for targeted suicide prevention interventions and improve research on the pathways from ideation to attempt.

A Study of Social and Behavioral Determinants of Health in Lung Cancer Patients Using Transformers-based Natural Language Processing Models.

Social and behavioral determinants of health (SBDoH) have important roles in shaping people's health. In clinical research studies, especially comparative effectiveness studies, failure to adjust for SBDoH factors will potentially cause confounding issues and misclassification errors in either statistical analyses and machine learning-based models. However, there are limited studies to examine SBDoH factors in clinical outcomes due to the lack of structured SBDoH information in current electronic health record (EHR) systems, while much of the SBDoH information is documented in clinical narratives. Natural language processing (NLP) is thus the key technology to extract such information from unstructured clinical text. However, there is not a mature clinical NLP system focusing on SBDoH. In this study, we examined two state-of-the-art transformer-based NLP models, including BERT and RoBERTa, to extract SBDoH concepts from clinical narratives, applied the best performing model to extract SBDoH concepts on a lung cancer screening patient cohort, and examined the difference of SBDoH information between NLP extracted results and structured EHRs (SBDoH information captured in standard vocabularies such as the International Classification of Diseases codes). The experimental results show that the BERT-based NLP model achieved the best strict/lenient F1-score of 0.8791 and 0.8999, respectively. The comparison between NLP extracted SBDoH information and structured EHRs in the lung cancer patient cohort of 864 patients with 161,933 various types of clinical notes showed that much more detailed information about smoking, education, and employment were only captured in clinical narratives and that it is necessary to use both clinical narratives and structured EHRs to construct a more complete picture of patients' SBDoH factors.

Toward cross-platform electronic health record-driven phenotyping using Clinical Quality Language.

INTRODUCTION: Electronic health record (EHR)-driven phenotyping is a critical first step in generating biomedical knowledge from EHR data. Despite recent progress, current phenotyping approaches are manual, time-consuming, error-prone, and platform-specific. This results in duplication of effort and highly variable results across systems and institutions, and is not scalable or portable. In this work, we investigate how the nascent Clinical Quality Language (CQL) can address these issues and enable high-throughput, cross-platform phenotyping. METHODS: We selected a clinically validated heart failure (HF) phenotype definition and translated it into CQL, then developed a CQL execution engine to integrate with the Observational Health Data Sciences and Informatics (OHDSI) platform. We executed the phenotype definition at two large academic medical centers, Northwestern Medicine and Weill Cornell Medicine, and conducted results verification (n = 100) to determine precision and recall. We additionally executed the same phenotype definition against two different data platforms, OHDSI and Fast Healthcare Interoperability Resources (FHIR), using the same underlying dataset and compared the results. RESULTS: CQL is expressive enough to represent the HF phenotype definition, including Boolean and aggregate operators, and temporal relationships between data elements. The language design also enabled the implementation of a custom execution engine with relative ease, and results verification at both sites revealed that precision and recall were both 100%. Cross-platform execution resulted in identical patient cohorts generated by both data platforms. CONCLUSIONS: CQL supports the representation of arbitrarily complex phenotype definitions, and our execution engine implementation demonstrated cross-platform execution against two widely used clinical data platforms. The language thus has the potential to help address current limitations with portability in EHR-driven phenotyping and scale in learning health systems.

Subphenotyping depression using machine learning and electronic health records.

OBJECTIVE: To identify depression subphenotypes from Electronic Health Records (EHRs) using machine learning methods, and analyze their characteristics with respect to patient demographics, comorbidities, and medications. MATERIALS AND METHODS: Using EHRs from the INSIGHT Clinical Research Network (CRN) database, multiple machine learning (ML) algorithms were applied to analyze 11 275 patients with depression to discern depression subphenotypes with distinct characteristics. RESULTS: Using the computational approaches, we derived three depression subphenotypes: Phenotype_A (n = 2791; 31.35%) included patients who were the oldest (mean (SD) age, 72.55 (14.93) years), had the most comorbidities, and took the most medications. The most common comorbidities in this cluster of patients were hyperlipidemia, hypertension, and diabetes. Phenotype_B (mean (SD) age, 68.44 (19.09) years) was the largest cluster (n = 4687; 52.65%), and included patients suffering from moderate loss of body function. Asthma, fibromyalgia, and Chronic Pain and Fatigue (CPF) were common comorbidities in this subphenotype. Phenotype_C (n = 1452; 16.31%) included patients who were younger (mean (SD) age, 63.47 (18.81) years), had the fewest comorbidities, and took fewer medications. Anxiety and tobacco use were common comorbidities in this subphenotype. CONCLUSION: Computationally deriving depression subtypes can provide meaningful insights and improve understanding of depression as a heterogeneous disorder. Further investigation is needed to assess the utility of these derived phenotypes to inform clinical trial design and interpretation in routine patient care.

Data-driven discovery of probable Alzheimer's disease and related dementia subphenotypes using electronic health records.

INTRODUCTION: We sought to assess longitudinal electronic health records (EHRs) using machine learning (ML) methods to computationally derive probable Alzheimer's Disease (AD) and related dementia subphenotypes. METHODS: A retrospective analysis of EHR data from a cohort of 7587 patients seen at a large, multi-specialty urban academic medical center in New York was conducted. Subphenotypes were derived using hierarchical clustering from 792 probable AD patients (cases) who had received at least one diagnosis of AD using their clinical data. The other 6795 patients, labeled as controls, were matched on age and gender with the cases and randomly selected in the ratio of 9:1. Prediction models with multiple ML algorithms were trained on this cohort using 5-fold cross-validation. XGBoost was used to rank the variable importance. RESULTS: Four subphenotypes were computationally derived. Subphenotype A (n = 273; 28.2%) had more patients with cardiovascular diseases; subphenotype B (n = 221; 27.9%) had more patients with mental health illnesses, such as depression and anxiety; patients in subphenotype C (n = 183; 23.1%) were overall older (mean (SD) age, 79.5 (5.4) years) and had the most comorbidities including diabetes, cardiovascular diseases, and mental health disorders; and subphenotype D (n = 115; 14.5%) included patients who took anti-dementia drugs and had sensory problems, such as deafness and hearing impairment.The 0-year prediction model for AD risk achieved an area under the receiver operating curve (AUC) of 0.764 (SD: 0.02); the 6-month model, 0.751 (SD: 0.02); the 1-year model, 0.752 (SD: 0.02); the 2-year model, 0.749 (SD: 0.03); and the 3-year model, 0.735 (SD: 0.03), respectively. Based on variable importance, the top-ranked comorbidities included depression, stroke/transient ischemic attack, hypertension, anxiety, mobility impairments, and atrial fibrillation. The top-ranked medications included anti-dementia drugs, antipsychotics, antiepileptics, and antidepressants. CONCLUSIONS: Four subphenotypes were computationally derived that correlated with cardiovascular diseases and mental health illnesses. ML algorithms based on patient demographics, diagnosis, and treatment demonstrated promising results in predicting the risk of developing AD at different time points across an individual's lifespan.

Identifying sub-phenotypes of acute kidney injury using structured and unstructured electronic health record data with memory networks.

Acute Kidney Injury (AKI) is a common clinical syndrome characterized by the rapid loss of kidney excretory function, which aggravates the clinical severity of other diseases in a large number of hospitalized patients. Accurate early prediction of AKI can enable in-time interventions and treatments. However, AKI is highly heterogeneous, thus identification of AKI sub-phenotypes can lead to an improved understanding of the disease pathophysiology and development of more targeted clinical interventions. This study used a memory network-based deep learning approach to discover AKI sub-phenotypes using structured and unstructured electronic health record (EHR) data of patients before AKI diagnosis. We leveraged a real world critical care EHR corpus including 37,486 ICU stays. Our approach identified three distinct sub-phenotypes: sub-phenotype I is with an average age of 63.03±17.25 years, and is characterized by mild loss of kidney excretory function (Serum Creatinine (SCr) 1.55±0.34 mg/dL, estimated Glomerular Filtration Rate Test (eGFR) 107.65±54.98 mL/min/1.73 m2). These patients are more likely to develop stage I AKI. Sub-phenotype II is with average age 66.81±10.43 years, and was characterized by severe loss of kidney excretory function (SCr 1.96±0.49 mg/dL, eGFR 82.19±55.92 mL/min/1.73 m2). These patients are more likely to develop stage III AKI. Sub-phenotype III is with average age 65.07±11.32 years, and was characterized moderate loss of kidney excretory function and thus more likely to develop stage II AKI (SCr 1.69±0.32 mg/dL, eGFR 93.97±56.53 mL/min/1.73 m2). Both SCr and eGFR are significantly different across the three sub-phenotypes with statistical testing plus postdoc analysis, and the conclusion still holds after age adjustment.

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.