JMIR Publications

ABSTRACT

Background:

The global cancer burden is rapidly increasing, with 20 million new cases estimated in 2022. The world population aged ≥65 years is also increasing, projected to reach 15.9% by 2050, making cancer control for older patients urgent. While surgical resection is key in cancer treatment, predicting postoperative disability and mortality in older patients with cancer is crucial for surgical decision-making, considering the quality of life and care burden. Currently, no model directly predicts postoperative functional disability in this population.

Objective:

We aimed to develop and validate machine-learning models for predicting postoperative functional disability (≥5-point decrease in Barthel Index) or in-hospital death in patients with cancer aged ≥65 years.

Methods:

This retrospective cohort study included patients aged ≥65 years who underwent surgery for major cancers (lung, stomach, colorectal, liver, pancreatic, breast, or prostate) in 70 Japanese hospitals. Patients were divided into development (April 2016–March 2022) and validation (April 2022–March 2023) cohorts. Before modeling, we selected predictor candidates with a crude odds ratio (OR) P-value of <.1, based on 37 routinely available preoperative factors through electronic medical records (e.g., age, sex, income, comorbidities, laboratory test values, and vital signs). Machine-learning models were developed using the least absolute shrinkage and selection operator (Lasso) logistic regression and random forest model algorithms. Model predictive performance was measured using the area under the receiver operating characteristic curve (AUC) with 95% confidence intervals (95%CI).

Results:

The study included 23,291 patients in the development cohort and 4,764 patients in the validation cohort. Based on the crude OR, 26 predictors were selected. Thereafter, the Lasso and random forest models used 15 and 26 predictors, respectively. Comparing the top 15 influential factors in the random forest model, we found 14 common factors in the Lasso model, including age ≥85 years (OR: 4.16, 95%CI: 3.37–5.14), dementia (OR: 3.84, 95%CI: 3.21–4.57), and non-home admission (OR: 3.05, 95%CI: 2.11–4.35). AUCs for the Lasso and random forest models were 0.78 (95%CI: 0.76–0.79) and 0.81 (95%CI: 0.80–0.83), respectively, in the development cohort. In the validation cohort, the AUCs for the Lasso and random forest models were comparable [0.78 (95%CI: 0.74–0.81) vs. 0.77 (95%CI: 0.74–0.81), respectively; P =.33].

Conclusions:

Both models demonstrated good performance in predicting postoperative outcomes for older patients with cancer, offering potential implementation in clinical settings through electronic medical record systems. They may support surgical decision-making for patients and families, and guide targeted interventions for healthcare providers. The Lasso model, with fewer predictors, may be preferable in time-sensitive clinical settings such as emergency surgery.