OptLung: An online predictive model for optimizing therapeutic strategies for second primary lung cancer in cancer survivors

Abstract

Background

Advances in cancer detection and treatment have significantly improved survival rates and led to an increasing prevalence of the second primary lung malignancy among cancer survivors. The prognosis and optimal treatment strategies of second primary lung malignancy differ substantially from those of first primary lung cancers. Existing prognostic models primarily focus on the first primary lung cancers and offer limited guidance for personalized treatment strategies to patients with second primary lung malignancy.

Methods

We identified patients with second primary lung malignancy who met the inclusion criteria in the Surveillance, Epidemiology, and End Results database. Machine learning models, including classification trees, K-nearest neighbors, gradient boosting machine, neural network, and random forest, were developed to predict 1- to 5-year overall survival and cancer-specific survival. The dataset was split into training and testing sets (8:2 ratio), and the model performance was evaluated using area under the curve, accuracy, sensitivity, specificity, precision, F1 scores, and Brier scores. By comparing survival outcomes between the concordant group (patients whose treatments aligned with model recommendations) and the discordant group (patients whose treatments did not), the ability of models to recommend optimal therapeutic strategy was validated.

Results

Among 32,370 patients with second primary lung malignancy from the Surveillance, Epidemiology, and End Results database, factors associated with worse prognosis included older age, male sex, White race, unmarried status, nonadenocarcinoma histology, advanced T and N stages, and advanced American Joint Committee on Cancer stage. Conversely, smaller tumor size, surgical intervention, chemotherapy, and radiotherapy were associated with improved prognosis. The gradient boosting machine model exhibited superior predictive performance for overall survival and cancer-specific survival and achieved area under the curve values exceeding 0.84 across all time points. Therapeutic recommendations from models proved effective because the concordant group demonstrated significantly better overall survival and cancer-specific survival than the discordant group. To enhance the clinical applicability of treatment recommendations by machine-learning models, an interactive web-based tool, OptLung (https://hznuduan.shinyapps.io/OptLung/), was developed.

Conclusion

This study used machine-learning models to accurately predict the survival curve of patients with second primary lung malignancy. These models built a user-friendly web-based platform where clinicians can obtain the optimal therapeutic strategy for patients with second primary lung malignancy.