Cindy A Crump1, Christian Wernz2,3, Loretta Schlachta-Fairchild1, Ernest Steidle1, Angela Duncan4, Lauretta Cathers1. 1. Doctoral Program in Health-Related Sciences, College of Health Professions, Virginia Commonwealth University, Richmond, Virginia, USA. 2. Department of Health Administration, College of Health Professions, Virginia Commonwealth University, Richmond, Virginia, USA. 3. Department of Statistical Sciences and Operations Research, College of Health Professions, Virginia Commonwealth University, Richmond, Virginia, USA. 4. Department of Patient Counseling, College of Health Professions, Virginia Commonwealth University, Richmond, Virginia, USA.
Abstract
Background: Clinical studies of telemedicine (TM) programs for chronic illness have demonstrated mixed results across settings and populations. With recent uptake in use of digital health modalities, more precise patient classification may improve outcomes, efficiency, and effectiveness. Objective: The purpose of the research was to develop a predictive score that measures the influence of patient characteristics on TM interventions. The central hypothesis is that disease type, illness severity, and the social determinants of health influence outcomes, including resource utilization, and can be precisely characterized. Methods: The retrospective study evaluated the feasibility of creating a patient "Telemedicine ImPact" (TIP) score derived from a Virginia Medicare and Medicaid claims data set. Claims were randomly selected, stratified by disease type, and matched by illness severity into a TM intervention group (N = 7,782) and a nontelemedicine "usual care" control cohort (N = 7,981). The individual records were then summarized into 15,762 cases with 80% of the cases used to develop, train, and test four predictive models (hospital utilization, readmissions, total utilization, and mortality) using 10-fold cross-validation. Results: Bayesian supervised machine learning achieved reference model performance index area under the curve for receiver operating characteristic (AUC/ROC) ≥0.85. Posterior probabilities for each outcome model were generated on a "hold-back" set of 3,082 cases. Robust parametric statistical methods enabled dimension reduction, model validation, and derivation of a reliable composite scaled score that quantified the overall health risk for each case. The TM intervention cohort demonstrated higher total utilization (representing the sum of inpatient, outpatient, and prescription use) and lower mean inpatient utilization than the usual standard of care. This finding suggests TM-based care may shift the composition of health resource utilization, reducing hospitalizations while increasing outpatient services, adjusted for patient differences. Conclusions: The creation of a patient score using machine learning to predict the effect of TM on outcomes is feasible. Adoption of the TIP score may reduce variability in results by more precisely accounting for the effects of patient characteristics on health outcomes and utilization. More consistent outcome prediction may lead to greater support for digital health.
Background: Clinical studies of telemedicine (TM) programs for chronic illness have demonstrated mixed results across settings and populations. With recent uptake in use of digital health modalities, more precise patient classification may improve outcomes, efficiency, and effectiveness. Objective: The purpose of the research was to develop a predictive score that measures the influence of patient characteristics on TM interventions. The central hypothesis is that disease type, illness severity, and the social determinants of health influence outcomes, including resource utilization, and can be precisely characterized. Methods: The retrospective study evaluated the feasibility of creating a patient "Telemedicine ImPact" (TIP) score derived from a Virginia Medicare and Medicaid claims data set. Claims were randomly selected, stratified by disease type, and matched by illness severity into a TM intervention group (N = 7,782) and a nontelemedicine "usual care" control cohort (N = 7,981). The individual records were then summarized into 15,762 cases with 80% of the cases used to develop, train, and test four predictive models (hospital utilization, readmissions, total utilization, and mortality) using 10-fold cross-validation. Results: Bayesian supervised machine learning achieved reference model performance index area under the curve for receiver operating characteristic (AUC/ROC) ≥0.85. Posterior probabilities for each outcome model were generated on a "hold-back" set of 3,082 cases. Robust parametric statistical methods enabled dimension reduction, model validation, and derivation of a reliable composite scaled score that quantified the overall health risk for each case. The TM intervention cohort demonstrated higher total utilization (representing the sum of inpatient, outpatient, and prescription use) and lower mean inpatient utilization than the usual standard of care. This finding suggests TM-based care may shift the composition of health resource utilization, reducing hospitalizations while increasing outpatient services, adjusted for patient differences. Conclusions: The creation of a patient score using machine learning to predict the effect of TM on outcomes is feasible. Adoption of the TIP score may reduce variability in results by more precisely accounting for the effects of patient characteristics on health outcomes and utilization. More consistent outcome prediction may lead to greater support for digital health.