AI-Driven Healthcare Management and Service Quality: Examining the Effects of Digital Transformation on Patient Satisfaction — A SimulationBased Framework Linking Clinic Operations to SERVQUAL Perceptions

Abstract

Hospitals are increasingly turning to AI for appointment scheduling, patient triage and flow and are anticipating more positive outcomes for satisfaction. Will this happen will depend on the impact that changes in operation have on the perceived service quality by patients. This paper proposes and implements a two-layer model, which connects them. The operational layer is a discrete-event simulation of an outpatient clinic executed as a 2³ leveled factorial experiment consisting of 500 simulated days (per scenario) using common random numbers. It assesses 3 different AI interventions: no show prediction and targeted reminders (NS), AI enabled triage (TR), and digital check in with pre-visit intake (DC). The simulated waiting times are transformed into top-box satisfaction through a response model with Probabilistic Sensitivity Analysis. The perceptual layer is a SERVLOAD based survey model with PLS estimated and consistent PLS (PLSc). All its inference properties are based on the Monte Carlo simulation of synthetic respondents; no patient data are used or claimed. NS increased mean waiting time by 5.5 min while delivering 10.0 additional patients per session, which was partially mitigated by the NS×DC interaction (−1.32 min), and by DC (−4.5 min). Consequently, the combined system's satisfaction level was located close to zero (+0.3 percentage points; 95% interval between minus 2.1 and plus 2.8), and DC alone increased the satisfaction level of 98% of the draws. TR was significant only at 115% load where it reduced wait time for urgent patients by 2.5 min. Group-wise reminder quota and assisted check-in closed the gap to patients' satisfaction at +1.2 points, but a less accurate model for patients with low digital literacy did so. Within the survey layer, PLS was 9% off the truth for the estimates of the paths, while PLSc was almost unbiased. These small hypothesized effects (β≠¼0.10) required several hundred or more than 1,000 respondents to reach 80% power, far more than are commonly believed to be necessary by the rules of thumb. The study concludes with design guidance for AI based clinics and a patient survey analysis protocol which is ready for use in future surveys.