Leveraging electronic health records for data science: common pitfalls and how to avoid them.

Analysis of electronic health records (EHRs) is an increasingly common approach for studying real-world patient data. Use of routinely collected data offers several advantages compared with other study designs, including reduced administrative costs, the ability to update analysis as practice patterns evolve, and larger sample sizes. Methodologically, EHR analysis is subject to distinct challenges because data are not collected for research purposes. In this Viewpoint, we elaborate on the importance of in-depth knowledge of clinical workflows and describe six potential pitfalls to be avoided when working with EHR data, drawing on examples from the literature and our experience. We propose solutions for prevention or mitigation of factors associated with each of these six pitfalls-sample selection bias, imprecise variable definitions, limitations to deployment, variable measurement frequency, subjective treatment allocation, and model overfitting. Ultimately, we hope that this Viewpoint will guide researchers to further improve the methodological robustness of EHR analysis.

Intraoperative prediction of postanaesthesia care unit hypotension.

BACKGROUND: Postoperative hypotension is associated with adverse outcomes, but intraoperative prediction of postanaesthesia care unit (PACU) hypotension is not routine in anaesthesiology workflow. Although machine learning models may support clinician prediction of PACU hypotension, clinician acceptance of prediction models is poorly understood. METHODS: We developed a clinically informed gradient boosting machine learning model using preoperative and intraoperative data from 88 446 surgical patients from 2015 to 2019. Nine anaesthesiologists each made 192 predictions of PACU hypotension using a web-based visualisation tool with and without input from the machine learning model. Questionnaires and interviews were analysed using thematic content analysis for model acceptance by anaesthesiologists. RESULTS: The model predicted PACU hypotension in 17 029 patients (area under the receiver operating characteristic [AUROC] 0.82 [95% confidence interval {CI}: 0.81-0.83] and average precision 0.40 [95% CI: 0.38-0.42]). On a random representative subset of 192 cases, anaesthesiologist performance improved from AUROC 0.67 (95% CI: 0.60-0.73) to AUROC 0.74 (95% CI: 0.68-0.79) with model predictions and information on risk factors. Anaesthesiologists perceived more value and expressed trust in the prediction model for prospective planning, informing PACU handoffs, and drawing attention to unexpected cases of PACU hypotension, but they doubted the model when predictions and associated features were not aligned with clinical judgement. Anaesthesiologists expressed interest in patient-specific thresholds for defining and treating postoperative hypotension. CONCLUSIONS: The ability of anaesthesiologists to predict PACU hypotension was improved by exposure to machine learning model predictions. Clinicians acknowledged value and trust in machine learning technology. Increasing familiarity with clinical use of model predictions is needed for effective integration into perioperative workflows.

Sociodemographic and clinical features predictive of SARS-CoV-2 test positivity across healthcare visit-types.

BACKGROUND: Despite increased testing efforts and the deployment of vaccines, COVID-19 cases and death toll continue to rise at record rates. Health systems routinely collect clinical and non-clinical information in electronic health records (EHR), yet little is known about how the minimal or intermediate spectra of EHR data can be leveraged to characterize patient SARS-CoV-2 pretest probability in support of interventional strategies. METHODS AND FINDINGS: We modeled patient pretest probability for SARS-CoV-2 test positivity and determined which features were contributing to the prediction and relative to patients triaged in inpatient, outpatient, and telehealth/drive-up visit-types. Data from the University of Washington (UW) Medicine Health System, which excluded UW Medicine care providers, included patients predominately residing in the Seattle Puget Sound area, were used to develop a gradient-boosting decision tree (GBDT) model. Patients were included if they had at least one visit prior to initial SARS-CoV-2 RT-PCR testing between January 01, 2020 through August 7, 2020. Model performance assessments used area-under-the-receiver-operating-characteristic (AUROC) and area-under-the-precision-recall (AUPR) curves. Feature performance assessments used SHapley Additive exPlanations (SHAP) values. The generalized pretest probability model using all available features achieved high overall discriminative performance (AUROC, 0.82). Performance among inpatients (AUROC, 0.86) was higher than telehealth/drive-up testing (AUROC, 0.81) or outpatient testing (AUROC, 0.76). The two-week test positivity rate in patient ZIP code was the most informative feature towards test positivity across visit-types. Geographic and sociodemographic factors were more important predictors of SARS-CoV-2 positivity than individual clinical characteristics. CONCLUSIONS: Recent geographic and sociodemographic factors, routinely collected in EHR though not routinely considered in clinical care, are the strongest predictors of initial SARS-CoV-2 test result. These findings were consistent across visit types, informing our understanding of individual SARS-CoV-2 risk factors with implications for deployment of testing, outreach, and population-level prevention efforts.

Early prediction of circulatory failure in the intensive care unit using machine learning.

Intensive-care clinicians are presented with large quantities of measurements from multiple monitoring systems. The limited ability of humans to process complex information hinders early recognition of patient deterioration, and high numbers of monitoring alarms lead to alarm fatigue. We used machine learning to develop an early-warning system that integrates measurements from multiple organ systems using a high-resolution database with 240 patient-years of data. It predicts 90% of circulatory-failure events in the test set, with 82% identified more than 2 h in advance, resulting in an area under the receiver operating characteristic curve of 0.94 and an area under the precision-recall curve of 0.63. On average, the system raises 0.05 alarms per patient and hour. The model was externally validated in an independent patient cohort. Our model provides early identification of patients at risk for circulatory failure with a much lower false-alarm rate than conventional threshold-based systems.

Identification of active transcriptional regulatory elements from GRO-seq data.

Modifications to the global run-on and sequencing (GRO-seq) protocol that enrich for 5'-capped RNAs can be used to reveal active transcriptional regulatory elements (TREs) with high accuracy. Here, we introduce discriminative regulatory-element detection from GRO-seq (dREG), a sensitive machine learning method that uses support vector regression to identify active TREs from GRO-seq data without requiring cap-based enrichment (https://github.com/Danko-Lab/dREG/). This approach allows TREs to be assayed together with gene expression levels and other transcriptional features in a single experiment. Predicted TREs are more enriched for several marks of transcriptional activation­including expression quantitative trait loci, disease-associated polymorphisms, acetylated histone 3 lysine 27 (H3K27ac) and transcription factor binding­than those identified by alternative functional assays. Using dREG, we surveyed TREs in eight human cell types and provide new insights into global patterns of TRE function.