Development of a Machine Learning-Based Patient Classification System for a Gastroenterology Ward

Patient classification systems (PCSs) are frequently used to estimate hours of nursing care to support budgeting and guide staffing decisions. However, despite a clear clinical need, there is as yet no validated and reliable PCS designed specifically for gastroenterology wards.

To develop a PCS tailored to gastroenterology wards that can accurately quantify nursing workload and support evidence-based nurse staffing decisions.

This study used Orem's self-care theory and Henderson's human needs theory as a foundation and applied machine learning techniques to construct a PCS for a gastroenterology ward in a general tertiary hospital. Patients were enrolled using convenience sampling, and information on their demographic characteristics, together with daily nursing activities and frequencies, was retrospectively extracted from the hospital information system (HIS) between July 1, 2019, and June 30, 2020. A workload measurement method was used to calculate the amount of nursing care received by each patient over a 24-hour period, thereby forming the study database. A decision tree model was developed to classify patients. The PCS was subsequently refined and validated using a prospective observational study of 357 patients conducted from December 1, 2022, to March 31, 2023.

The final PCS included two primary categories and five subcategories, each of which was defined by specific patient characteristics and their corresponding 24-hour nursing time requirements: Category 1: "Day of surgery and first postoperative day," in which patients were subdivided into three levels based on surgical complexity, self-care capacity, and illness severity (Surgery 1, Surgery 2, and Surgery 3), requiring 1.66, 2.82, and 4.15 nursing hours per 24 h, respectively, and Category 2: "Other days" patients, who were subdivided into two groups based on critical illness, self-care ability, and disease severity (Category 1 and Category 2), with each requiring 0.96 and 3.42 nursing hours per 24 h, respectively. Internal and external validation demonstrated good model fit and acceptable predictive performance.

Patients could be rapidly classified using defined indicators, enabling accurate prediction of their 24-hour nursing care requirements. The PCS exhibited strong internal consistency, stability, and generalizability, thereby supporting the reliability of its results.

This PCS provides a scientific basis for nurse staffing decisions and may support hospital administrators and health authorities in the development of data-driven nurse workforce allocation policies.