Application of unsupervised deep learning algorithms for identification of specific clusters of chronic cough patients from EMR data.

BACKGROUND: Chronic cough affects approximately 10% of adults. The lack of ICD codes for chronic cough makes it challenging to apply supervised learning methods to predict the characteristics of chronic cough patients, thereby requiring the identification of chronic cough patients by other mechanisms. We developed a deep clustering algorithm with auto-encoder embedding (DCAE) to identify clusters of chronic cough patients based on data from a large cohort of 264,146 patients from the Electronic Medical Records (EMR) system. We constructed features using the diagnosis within the EMR, then built a clustering-oriented loss function directly on embedded features of the deep autoencoder to jointly perform feature refinement and cluster assignment. Lastly, we performed statistical analysis on the identified clusters to characterize the chronic cough patients compared to the non-chronic cough patients. RESULTS: The experimental results show that the DCAE model generated three chronic cough clusters and one non-chronic cough patient cluster. We found various diagnoses, medications, and lab tests highly associated with chronic cough patients by comparing the chronic cough cluster with the non-chronic cough cluster. Comparison of chronic cough clusters demonstrated that certain combinations of medications and diagnoses characterize some chronic cough clusters. CONCLUSIONS: To the best of our knowledge, this study is the first to test the potential of unsupervised deep learning methods for chronic cough investigation, which also shows a great advantage over existing algorithms for patient data clustering.

Development, validation, and proof-of-concept implementation of a two-year risk prediction model for undiagnosed atrial fibrillation using common electronic health data (UNAFIED).

BACKGROUND: Many patients with atrial fibrillation (AF) remain undiagnosed despite availability of interventions to reduce stroke risk. Predictive models to date are limited by data requirements and theoretical usage. We aimed to develop a model for predicting the 2-year probability of AF diagnosis and implement it as proof-of-concept (POC) in a production electronic health record (EHR). METHODS: We used a nested case-control design using data from the Indiana Network for Patient Care. The development cohort came from 2016 to 2017 (outcome period) and 2014 to 2015 (baseline). A separate validation cohort used outcome and baseline periods shifted 2 years before respective development cohort times. Machine learning approaches were used to build predictive model. Patients ≥ 18 years, later restricted to age ≥ 40 years, with at least two encounters and no AF during baseline, were included. In the 6-week EHR prospective pilot, the model was silently implemented in the production system at a large safety-net urban hospital. Three new and two previous logistic regression models were evaluated using receiver-operating characteristics. Number, characteristics, and CHA2DS2-VASc scores of patients identified by the model in the pilot are presented. RESULTS: After restricting age to ≥ 40 years, 31,474 AF cases (mean age, 71.5 years; female 49%) and 22,078 controls (mean age, 59.5 years; female 61%) comprised the development cohort. A 10-variable model using age, acute heart disease, albumin, body mass index, chronic obstructive pulmonary disease, gender, heart failure, insurance, kidney disease, and shock yielded the best performance (C-statistic, 0.80 [95% CI 0.79-0.80]). The model performed well in the validation cohort (C-statistic, 0.81 [95% CI 0.8-0.81]). In the EHR pilot, 7916/22,272 (35.5%; mean age, 66 years; female 50%) were identified as higher risk for AF; 5582 (70%) had CHA2DS2-VASc score ≥ 2. CONCLUSIONS: Using variables commonly available in the EHR, we created a predictive model to identify 2-year risk of developing AF in those previously without diagnosed AF. Successful POC implementation of the model in an EHR provided a practical strategy to identify patients who may benefit from interventions to reduce their stroke risk.

A Feasibility Study: Testing Whether a Sleep Application Providing Objective Sleep Data to Physicians Improves Patient-Physician Communication Regarding Sleep Experiences, Habits, and Behaviors.

INTRODUCTION: Sleep tracker data have not been utilized routinely in sleep-related disorders and their management. Sleep-related disorders are common in primary care practice and incorporating sleep tracker data may help in improving patient care. We conducted a pilot study to assess the feasibility of a sleep program using the Fitbit Charge 2™ device and SleepLife® application. The main aim of the study was to examine whether a program using a commercially available wearable sleep tracker device providing objective sleep data would improve communication in primary care settings between patients and their providers. Secondary aims included whether patient satisfaction with care would improve as result of the program. METHODS: A prospective, randomized, parallel group, observational pilot study was conducted in 20 primary care clinics in Indianapolis, IN from June 2018 to February 2019. Inclusion criteria included patients over the age of 18, have a diagnosis of insomnia identified by electronic medical record and/or a validated questionnaire, and were on a prescription sleep aid. The study was not specific to any sleep aid prescription, branded or generic, and was not designed to evaluate a drug or drug class. Each primary care clinic was randomized to either the SleepLife® intervention or the control arm. All patients were provided with a Fitbit Charge 2™ device. Only patients in the intervention arm were educated on how to use the SleepLife® application. Physicians in the intervention arm were set up with the SleepLife® portal on their computers. RESULTS: Forty-nine physicians and 75 patients were enrolled in the study. Patients had a mean age of 57 (SD 12.8) years and 61% were female. Mean age of physicians was 47 (SD 10.6) years. Patients showed high rates of involvement in the program with 83% completing all survey questions. Physician survey completion rate was 55%. Only one physician logged into the SleepLife portal to check their patients' sleep status. At the end of the 6-week intervention, patients' composite general satisfaction scores with sleep health management decreased significantly in the intervention arm when compared to controls (p = 0.03). Patients' satisfaction with communication also decreased significantly in the intervention group (p = 0.01). The sleep outcomes, which were calculated on the basis of study questionnaire answers, improved significantly in the intervention group as compared to the control group (p = 0.04). Physician communication satisfaction scores remained unchanged (p = 0.12). CONCLUSIONS: SleepLife® and its related physician portal can facilitate physician-patient communication, and it captures patient sleep outcomes including behaviors and habits. Patients were highly engaged with the program, while physicians did not demonstrate engagement. The study design and questionnaires do not specifically address the reasons behind the decreased patient satisfaction with care and communication, but it was perceived to be a result of physician non-responsiveness. Sleep quality scores on the other hand showed an improvement among SleepLife® users, suggesting that patients may have implemented good sleep practices on their own. Given that it was a feasibility study, and the sample size was small, we were not able to make major inferences regarding the difference between sleep disorder types. Additionally, we excluded patients with a history of alcohol use, substance abuse, or depression because of concerns that they may affect sleep independently. To promote the growth of technology in primary care, further research incorporating results from this study and physician engagement techniques should be included.

Applying interpretable deep learning models to identify chronic cough patients using EHR data.

BACKGROUND AND OBJECTIVE: Chronic cough (CC) affects approximately 10% of adults. Many disease states are associated with chronic cough, such as asthma, upper airway cough syndrome, bronchitis, and gastroesophageal reflux disease. The lack of an ICD code specific for chronic cough makes it challenging to identify such patients from electronic health records (EHRs). For clinical and research purposes, computational methods using EHR data are urgently needed to identify chronic cough cases. This research aims to investigate the data representations and deep learning algorithms for chronic cough prediction. METHODS: Utilizing real-world EHR data from a large academic healthcare system from October 2005 to September 2015, we investigated Natural Language Representation of the EHR data and systematically evaluated deep learning and traditional machine learning models to predict chronic cough patients. We built these machine learning models using structured data (medication and diagnosis) and unstructured data (clinical notes). RESULTS: The sensitivity and specificity of a transformer-based deep learning algorithm, specifically BERT with attention model, was 0.856 and 0.866, respectively, using structured data (medication and diagnosis). Sensitivity and specificity improved to 0.952 and 0.930 when we combined structured data with symptoms extracted from clinical notes. We further found that the attention mechanism of deep learning models can be used to extract important features that drive the prediction decisions. Compared with our previously published rule-based algorithm, the deep learning algorithm can identify more chronic cough patients with structured data. CONCLUSIONS: By applying deep learning models, chronic cough patients can be reliably identified for prospective or retrospective research through medication and diagnosis data, widely available in EHR and electronic claims data, thus improving the generalizability of the patient identification algorithm. Deep learning models can identify chronic cough patients with even higher sensitivity and specificity when structured and unstructured EHR data are utilized. We anticipate language-based data representation and deep learning models developed in this research could also be productively used for other disease prediction and case identification.

Survival, Chemotherapy Treatments, and Health Care Utilization Among Patients with Advanced Small Cell Lung Cancer: An Observational Study.

INTRODUCTION: Most cases of small cell lung cancer (SCLC) are diagnosed at an advanced stage. The objective of this study was to investigate patient characteristics, survival, chemotherapy treatments, and health care use after a diagnosis of advanced SCLC in subjects enrolled in a health system network. METHODS: This was a retrospective cohort study of patients aged ≥ 18 years who either were diagnosed with stage III/IV SCLC or who progressed to advanced SCLC during the study period (2005-2015). Patients identified from the Indiana State Cancer Registry and the Indiana Network for Patient Care were followed from their advanced diagnosis index date until the earliest date of the last visit, death, or the end of the study period. Patient characteristics, survival, chemotherapy regimens, associated health care visits, and durations of treatment were reported. Time-to-event analyses were performed using the Kaplan-Meier method. RESULTS: A total of 498 patients with advanced SCLC were identified, of whom 429 were newly diagnosed with advanced disease and 69 progressed to advanced disease during the study period. Median survival from the index diagnosis date was 13.2 months. First-line (1L) chemotherapy was received by 464 (93.2%) patients, most commonly carboplatin/etoposide, received by 213 (45.9%) patients, followed by cisplatin/etoposide (20.7%). Ninety-five (20.5%) patients progressed to second-line (2L) chemotherapy, where topotecan monotherapy (20.0%) was the most common regimen, followed by carboplatin/etoposide (14.7%). Median survival was 10.1 months from 1L initiation and 7.7 months from 2L initiation. CONCLUSION: Patients in a regional health system network diagnosed with advanced SCLC were treated with chemotherapy regimens similar to those in earlier reports based on SEER-Medicare data. Survival of patients with advanced SCLC was poor, illustrating the lack of progress over several decades in the treatment of this lethal disease and highlighting the need for improved treatments.

A randomized study on the usefulness of an electronic outpatient hypoglycemia risk calculator for clinicians of patients with diabetes in a safety-net institution.

Objective: Hypoglycemia (HG) occurs in up to 60% of patients with diabetes mellitus (DM) each year. We assessed a HG alert tool in an electronic health record system, and determined its effect on clinical practice and outcomes.Methods: The tool applied a statistical model, yielding patient-specific information about HG risk. We randomized outpatient primary-care providers (PCPs) to see or not see the alerts. Patients were assigned to study group according to the first PCP seen during four months. We assessed prescriptions, testing, and HG. Variables were compared by multinomial, logistic, or linear model. ClinicalTrials.gov ID: NCT04177147 (registered on 22 November 2019).Results: Patients (N = 3350) visited 123 intervention PCPs; 3395 patients visited 220 control PCPs. Intervention PCPs were shown 18,645 alerts (mean of 152 per PCP). Patients' mean age was 55 years, with 61% female, 49% black, and 49% Medicaid recipients. Mean baseline A1c and body mass index were similar between groups. During follow-up, the number of A1c and glucose tests, and number of new, refilled, changed, or discontinued insulin prescriptions, were highest for patients with highest risk. Per 100 patients on average, the intervention group had fewer sulfonylurea refills (6 vs. 8; p < .05) and outpatient encounters (470 vs. 502; p < .05), though the change in encounters was not significant. Frequency of HG events was unchanged.Conclusions: Informing PCPs about risk of HG led to fewer sulfonylurea refills and visits. Longer-term studies are needed to assess potential for long-term benefits.

Quantifying Unmet Need in Statin-Treated Hyperlipidemia Patients and the Potential Benefit of Further LDL-C Reduction Through an EHR-Based Retrospective Cohort Study.

BACKGROUND: Statins are effective in helping prevent cardiovascular disease (CVD). However, studies suggest that only 20%-64% of patients taking statins achieve reasonable low-density lipoprotein cholesterol (LDL-C) thresholds. On-treatment levels of LDL-C remain a key predictor of residual CVD event risk. OBJECTIVES: To (a) determine how many patients on statins achieved the therapeutic threshold of LDL-C < 100 mg per dL (general cohort) and < 70 mg per dL (secondary prevention cohort, or subcohort, with preexisting CVD); (b) estimate the number of potentially avoidable CVD events if the threshold were reached; and (c) forecast potential cost savings. METHODS: A retrospective, longitudinal cohort study using electronic health record data from the Indiana Network for Patient Care (INPC) was conducted. The INPC provides comprehensive information about patients in Indiana across health care organizations and care settings. Patients were aged > 45 years and seen between January 1, 2012, and October 31, 2016 (ensuring study of contemporary practice), were statin-naive for 12 months before the index date of initiating statin therapy, and had an LDL-C value recorded 6-18 months after the index date. Subsequent to descriptive cohort analysis, the theoretical CVD risk reduction achievable by reaching the threshold was calculated using Framingham Risk Score and Cholesterol Treatment Trialists' Collaboration formulas. Estimated potential cost savings used published first-year costs of CVD events, adjusted for inflation and discounted to the present day. RESULTS: Of the 89,267 patients initiating statins, 30,083 (33.7%) did not achieve the LDL-C threshold (subcohort: 58.1%). In both groups, not achieving the threshold was associated with patients who were female, black, and those who had reduced medication adherence. Higher levels of preventive aspirin use and antihypertensive treatment were associated with threshold achievement. In both cohorts, approximately 64% of patients above the threshold were within 30 mg per dL of the respective threshold. Adherence to statin therapy regimen, judged by a medication possession ratio of ≥ 80%, was 57.4% in the general cohort and 56.7% in the subcohort. Of the patients who adhered to therapy, 23.7% of the general cohort and 50.5% of the subcohort had LDL-C levels that did not meet the threshold. 10-year CVD event risk in the at-or-above threshold group was 22.78% (SD = 17.24%) in the general cohort and 29.56% (SD = 18.19%) in the subcohort. By reducing LDL-C to the threshold, a potential relative risk reduction of 14.8% in the general cohort could avoid 1,173 CVD events over 10 years (subcohort: 15.7% and 454 events). Given first-year inpatient and follow-up costs of $37,300 per CVD event, this risk reduction could save about $1,455 per patient treated to reach the threshold (subcohort: $1,902; 2017 U.S. dollars) over a 10-year period. CONCLUSIONS: Across multiple health care systems in Indiana, between 34% (general cohort) and 58% (secondary prevention cohort) of patients treated with statins did not achieve therapeutic LDL-C thresholds. Based on current CVD event risk and cost projections, such patients seem to be at increased risk and may represent an important and potentially preventable burden on health care costs. DISCLOSURES: Funding support for this study was provided by Merck (Kenilworth, NJ). Chase and Boggs are employed by Merck. Simpson is a consultant to Merck and Pfizer. The other authors have nothing to disclose.

Predictive modeling of hypoglycemia for clinical decision support in evaluating outpatients with diabetes mellitus.

Objective: Hypoglycemia occurs in 20-60% of patients with diabetes mellitus. Identifying at-risk patients can facilitate interventions to lower risk. We sought to develop a hypoglycemia prediction model. Methods: In this retrospective cohort study, urban adults prescribed a diabetes drug between 2004 and 2013 were identified. Demographic and clinical data were extracted from an electronic medical record (EMR). Laboratory tests, diagnostic codes and natural language processing (NLP) identified hypoglycemia. We compared multiple logistic regression, classification and regression trees (CART), and random forest. Models were evaluated on an independent test set or through cross-validation. Results: The 38,780 patients had mean age 57 years; 56% were female, 40% African-American and 39% uninsured. Hypoglycemia occurred in 8128 (539 identified only by NLP). In logistic regression, factors positively associated with hypoglycemia included infection, non-long-acting insulin, dementia and recent hypoglycemia. Negatively associated factors included long-acting insulin plus sulfonylurea, and age 75 or older. The models' area under curve was similar (logistic regression, 89%; CART, 88%; random forest, 90%, with ten-fold cross-validation). Conclusions: NLP improved identification of hypoglycemia. Non-long-acting insulin was an important risk factor. Decreased risk with age may reflect treatment or diminished awareness of hypoglycemia. More complex models did not improve prediction.