Abstract
OBJECTIVE: Patient-physician communication affects cancer patients' satisfaction, health outcomes, and reimbursement for physician services. Our objective is to use machine learning to comprehensively examine the association between patient satisfaction and physician factors in clinical consultations about cancer prognosis and pain. METHODS: We used data from audio-recorded, transcribed communications between physicians and standardized patients (SPs). We analyzed the data using logistic regression (LR) and random forests (RF). RESULTS: The LR models suggested that lower patient satisfaction was associated with more in-depth prognosis discussion; and higher patient satisfaction was associated with a greater extent of shared decision making, patient being black, and doctor being young. Conversely, the RF models suggested the opposite association with the same set of variables. CONCLUSION: Somewhat contradicting results from distinct machine learning models suggested possible confounding factors (hidden variables) in prognosis discussion, shared decision-making, and doctor age, on the modeling of patient satisfaction. Practitioners should not make inferences with one single data-modeling method and enlarge the study cohort to help deal with population heterogeneity. INNOVATION: Comparing diverse machine learning models (both parametric and non-parametric types) and carefully applying variable selection methods prior to regression modeling, can enrich the examination of physician factors in characterizing patient-physician communication outcomes.