Authors (including presenting author) :
Lee TH (1), Yu KK (1). Fong CH (1), Tommy Ng (2)
Affiliation :
(1) Department of rehabilitation, Kowloon Hospital (2) KCC IT department
Introduction :
Predicting the patient’s length of stay (LOS) in the hospital can facilitate early discharge through effective resource allocation. Clinical data of patients transferred to a regional rehabilitation hospital (Kowloon Hospital) were analyzed, and a prediction model was developed to answer the binary outcomes of non-LOS vs. long LOS in our rehabilitation unit.
Objectives :
To predict the LOS of patients in convalescent hospital
Methodology :
Two years of clinical data from CMS, including demographics, social background, diagnosis, physical function, and laboratory data, were retrieved. Random Forest (RF), Neural Network (NN), and Logistic Regression (LR) were used to predict the possibility of non-long LOS (≤ 21 days) vs long LOS (> 21 days)
Result & Outcome :
Results: From 1/7/21 to 30/6/23, 15190 patients were transferred from Queen Elizabeth Hospital (QEH) to KH for rehabilitation or convalescent care. 9957 patients (65.5%) were discharged within 21 days, while 5233 patients (34.5%) were hospitalized more than 21 days. Possible LOS predictors, including principal diagnosis, age, sex, acute hospital LOS, accommodation, household members, employment status, smoking status, mobility status, number of assistances on mobility, malnutrition screening tool score, and laboratory results were analyzed through various predictive models (RF, NN, LR). LOS in acute hospital was strongly correlated with LOS in convalescent hospital. Some predictors were found to be disease-specific. For example, the mean LOS in all old-aged home (OAH) patients was 19.1, which is no different from the non-OAH patients of 21.6 days. However, in stroke patients, the mean LOS in OAH patients was 21.5 days which was significantly shorter than LOS in non-OAH patients 27.8 days (p< 0.05). Random forest (RF) yielded the highest overall accuracy of 72.1% in predicting long LOS, with the positive predictive value of 66.8% (sensitivity = 32.7. %), the negative predictive value of 72.8% (specificity = 92.3%), the area under the curve (AUC) is 72.2% and F1 score 44.2. Conclusion: Analysis results from this LOS prediction model provide a fundament basis for future machine learning model development for AI prediction of patient LOS in the rehabilitation unit
