DEVELOPMENT OF A MACHINE LEARNING MODEL FOR FORECASTING HEALTHCARE SERVICE UTILIZATION BASED ON EMR DATA

Abstract

This study develops and evaluates a reproducible proof-of-concept (PoC) pipeline for forecasting day-ahead hospital service utilization from routinely collected electronic medical record (EMR) data. Three anonymized Excel extracts from August 2025 were used—outpatient visits (4,247 rows), laboratory actions (2,044 rows; August subset), and radiology examinations (31 rows)—then cleaned, time-parsed, and aggregated to daily counts. Features comprised day-of-week (DOW) and lagged outpatient volume (OPD_{t−1}) to avoid same-day leakage. We trained a linear regression model to predict daily laboratory counts and a Random Forest to flag daily high/low laboratory workload (threshold = within-month median ≈79 actions/day), evaluated with blocked time-series cross-validation. Outpatient visits ranged 154–204/day (median ≈172) and laboratory actions 61–171/day (median ≈79); radiology volumes were too sparse for reliable day-level modeling. The regression baseline achieved R² = −20.52 (mean CV), indicating that OPD_{t−1}+DOW alone cannot explain intra-month variability, while the classifier reached Accuracy = 0.53 ± 0.16, sufficient to validate the end-to-end pipeline but inadequate for autonomous staffing or reagent planning. These findings suggest that accurate operational forecasting requires multi-month horizons and clinically richer features (e.g., sending unit/poliklinik, grouped test panels, provider schedules, payer, and holiday effects) alongside additional temporal signals (multi-day lags/leads, rolling statistics).