A Machine Learning-Derived Risk Scorecard for Pneumonia Hospitalization in Japanese Old-Old Adults

AIM: To develop and internally validate a machine learning-based risk scorecard for 1-year pneumonia hospitalization among community-dwelling Japanese adults aged ≥ 75 years using routinely collected frailty screening and claims data. METHODS: We conducted a retrospective cohort study of 1 098 404 community-dwelling adults aged ≥ 75 years who completed the Questionnaire for Medical Checkup of Old-Old (QMCOO) between April 2020 and March 2024, using the Late-Stage Medical Care System claims database. Data were split into training (70%) and test (30%) sets. A Super Learner ensemble of 20 base learners was developed to predict 1-year pneumonia hospitalization (ICD-10: J12-J18, J69). An independent 17-feature point-based scorecard (0-21 points) was derived from the training set with Platt calibration. Performance was evaluated by the area under the receiver operating characteristic curve (AUC), calibration slope, and calibration-in-the-large. RESULTS: Among 1 098 404 participants (mean age 80.6 [SD 5.0] years; 40.3% male), 4525 (0.41%) experienced pneumonia hospitalization. On the test set, the Super Learner achieved an AUC of 0.823 (95% CI, 0.812-0.834) and the scorecard 0.786 (0.774-0.798), which showed good calibration (slope 1.017; calibration-in-the-large -0.001). Stratification into low (0-6 points; 55.6%, 0.12% event rate), moderate (7-10; 35.4%, 0.49%), and high (≥ 11; 9.0%, 1.91%) groups yielded a 15.9-fold risk gradient. CONCLUSIONS: A QMCOO-based risk scorecard showed good discrimination and calibration for predicting 1-year pneumonia hospitalization in this population. If externally validated, this tool may help identify high-risk individuals during routine health checkups and support targeted preventive assessment in primary care.