Optimizing chemotherapy decisions in low-resource settings with deep learning: A novel approach for unresectable NSCLC.

Abstract

e13668 Background: Even though chemotherapy offers less survival advantages in the era of targeted therapy and immunotherapy, nevertheless, it remains the cornerstone in low-resource settings. Here, we propose a novel approach based on deep learning to personalize chemotherapy regimens for patients with unresectable non-small cell lung cancer (NSCLC) to prolong progression-free survival (PFS) time. Methods: To enhance the practice and derive maximum benefit from chemotherapeutic agents to delay the time to progression, we developed a model based on a deep neural network (feature extraction) combined with a random forest (classification) that guides the choice of the chemotherapeutic agents used through a complete stepwise research process starting by incorporating various patient, disease, and toxicity factors and results in a suggestion of used regimens while manually backing the algorithm with new data allowing update. Sample data from 1265 patients were accurately collected from one Moroccan academic institute. The model underwent 10-fold cross-validation to assess its stability across different data subsets and was internally validated. The predictive power was evaluated using performance metrics such as F1 score and confusion matrix. Results: Twenty-four annotated categorical data from 882 patients were selected to build the model based on the results of the deep neural network, providing recommendations on where to place chemotherapy based on docetaxel, etoposide, paclitaxel, vinorelbine, gemcitabine, and pemetrexed, either in monotherapy or in combination with platinum. An effective learning process avoiding overfitting is demonstrated by the convergence of the training and validation loss curves, with a steady decrease in the training loss and a slight decrease in the validation loss, accompanied by fluctuations over 100 epochs. The evaluation was based on the results of unseen data, achieving an overall accuracy of 76.3% (95% Confidence Interval [CI], 71.3 to 81.06) with a Kappa value of 0.72. Class-wise performance showed the highest sensitivity of 86.06% for Class 3, demonstrating the model's strong ability to identify patient responses to specific chemotherapy regimens. The specificity for Class 1 was notably high at 99.72%, and the balanced accuracy for Class 7 was 87.44%, demonstrating robustness and reliability in providing personalized chemotherapy recommendations. Conclusions: Although the novelty of this research created with chemotherapeutic regimens is relatively limited in an era of advanced therapy, we demonstrated the potential of artificial intelligence (AI) in maximizing benefits for limited resources while accelerating scientific discovery in data-driven cancer research and beyond. We call for future collaboration to validate these findings on targeted therapy, immunotherapy, and especially antibody-drug conjugates.