Thorac Cancer. 2022 Oct 24. doi: 10.1111/1759-7714.14694. Online ahead of print.
BACKGROUND: The present study aimed to evaluate the performance of several machine learning (ML) algorithms in predicting 1-year afatinib continuation and 2-year survival after afatinib initiation and to identify the differences in survival outcomes between ML-classified strata.
METHODS: Data that were also used in the RESET study were retrospectively collected from 16 hospitals in South Korea. A stratified random sampling method was applied to split the data into training and test sets (70:30 split ratio). Clinical information, such as age, sex, tumor stage, smoking, performance status, metastasis, type of metastasis, dose adjustment, and pathologic information on EGFR mutations were inputted. Training was performed using eight ML algorithms: logistic regression, decision tree, deep neural network, random forest, support vector machine, boosting, bagging, and the naïve Bayes classifier. The model performance was assessed based on sensitivity, specificity, and accuracy. Area under the receiver operator characteristic curve (AUC) was calculated and compared between the ML models using DeLong’s test. A Kaplan-Meier (KM) curve was used to visualize the identified strata obtained from the ML models.
RESULTS: No significant differences in the input variables were observed between the training and test datasets. The best-performing models were support vector machine in predicting 1-year afatinib continuation (AUC 0.626) and decision tree in 2-year survival after afatinib start (AUC 0.644), although the performances of the ML models were comparable and did not display any predictive roles. KM analysis and log-rank test revealed significant differences between the strata identified from the ML model (p < 0.001) in terms of both time-on-treatment (TOT) and overall survival (OS).
CONCLUSION: The performances of ML models in our study found no discernible roles in predicting afatinib-related outcomes, although the identified strata revealed different TOT and OS in the KM analysis. This implies the strength of ML in predicting the survival outcome, as well as the limitation of electronic medical record-based variables in ML algorithms. Careful consideration of variable inclusion is likely to improve the general model performance.