Profiling unmet post-acute care needs of an inpatient population in Hong Kong: can real-world data and machine learning algorithms bring precision to tertiary prevention in the community?

Introduction: Case-mix systems aim to optimise acute care resource allocation, yet patients within the same groups often exhibit substantial variability in utilisation. This study aimed to examine how incorporating measures of clinical complexity and post-acute care utilisation-both critical to rehospitalisation risk and accurate resource planning-into case-mix stratification could improve the precision of acute care resource allocation.
Methods: Through iterative applications of unsupervised and supervised machine learning models, we extracted typical patient profiles from the study populations, analysed post-acute care utilisation patterns, and assessed the 28-day rehospitalisation rates resulting from different pairings between clinical profiles and post-acute care service utilisation patterns.
Results: Across various disease systems and age-groups, patients discharged without receiving algorithm-selected post-acute care (ie, No Service groups [NS groups]) showed significantly higher 28-day rehospitalisation rates relative to their corresponding segments in the same medoid case-mix groups (CMGs; pooled odds ratio [OR]=19.27; P<0.001). The NS groups also demonstrated higher rates of having two or more chronic diseases (pooled OR=1.84; P<0.001) and-for the 50-64-year-old population-resource-intensifying co-morbidities (pooled OR=1.23; P=0.05). Patients displaying higher rates of resource-intensifying co-morbidities compared with their ≥65-year-old counterparts (such as when the medoid CMG was renal failure or chronic obstructive pulmonary disease) also exhibited significantly higher 28-day rehospitalisation rates than the ≥65-year-old NS groups sharing the same medoid CMGs.
Conclusion: These findings support a precision-driven approach to designing rehospitalisation prevention programmes that target individuals aged 50 to 64 years discharged with specific clinical profiles, and developing and allocating human capital for these targeted prevention programmes.
© 2025 Hong Kong Academy of Medicine. CC BY-NC-ND 4.0

引言：病例組系統旨在優化急症護理資源分配，惟同一病例組內患者 的資源使用情況往往存在明顯差異。本研究旨在探討於病例組分層中 納入臨床複雜度及急性期後護理使用情況這兩項對再入院風險及資源 規劃準確度均具關鍵影響的指標，能否提高急症護理資源分配的精準 度。
方法：本研究透過反覆運用非監督及監督式機器學習模型，從研究群 組中擷取典型患者概況，分析急性期後護理服務的使用模式，並評估 不同臨床概況與急性期後護理服務使用組合所產生的28天再入院率。
結果：在不同疾病系統及年齡層中，出院時未獲演算法選定的急性期 後護理患者（即無服務組）的 28 天再入院率，顯著高於其所屬中位 病例組內相應分段的患者為高（合併比值比 = 19.27；P<0.001）。無 服務組中同時患有兩種或以上慢性疾病的比例亦較高（合併比值比 = 1.84；P<0.001），而在 50 至 64 歲群組中，無服務組出現高資源需 求共病的比率亦較其相應分段為高（合併比值比= 1.23；P=0.05）。 在若干中位病例組配對中（例如中位病例組為腎衰竭或慢性阻塞性肺 病時），50 至 64 歲無服務組患者不僅有較高的高資源需求共病比率， 其 28 天再入院率亦顯著高於同屬該病例組的65歲或以上無服務組患 者。
結論：本研究結果支持以精準為導向的策略，包括針對具有特定臨床 概況的 50 至 64 歲出院患者設計再入院預防計劃，以及按此等目標性預 防計劃的群組特徵培訓及分配所需的人力資源。