Predicting Chemotherapy-Induced Peripheral Neuropathy Using Transformer-Based Multimodal Deep Learning

Abstract

Background: Chemotherapy-induced peripheral neuropathy (CIPN) is a common and debilitating adverse effect of cancer treatment that substantially impairs patients’ quality of life and may lead to dose reduction or treatment discontinuation. Traditional prediction models based on single-modal data have shown limited accuracy in clinical settings.

Aim: This study aimed to develop and evaluate a deep learning-based predictive model for CIPN by integrating multimodal data, including clinical, genomic, biosignal, wearable device, and imaging information.

Methods: A retrospective and prospective cohort of cancer patients receiving chemotherapy between 2020 and 2025 was analyzed using data collected from multicenter electronic health records (EHRs) and public databases. An intermediate fusion framework was implemented using a Transformer-based architecture, which was compared with LSTM, CNN, and XGBoost models. SHAP (Shapley additive explanations) and Grad-CAM were used to improve model interpretability, while performance was assessed using AUC-ROC (area under the receiver operating characteristic curve), accuracy, sensitivity, specificity, and F1-score.

Results: The Transformer-based model achieved the highest performance (AUC = 0.93; accuracy = 88.5%; sensitivity = 85.3%; specificity = 90.1%), outperforming conventional models. SHAP analysis identified chemotherapy dosage, nerve magnetic resonance imaging abnormalities, electrocardiogram changes, CYP2C8 mutations, and diabetes as the most influential predictors. Patients with a high predicted risk of CIPN also demonstrated significantly lower overall survival, indicating a broader systemic impact of CIPN beyond neurological symptoms.

Conclusions: This study provides evidence that deep learning models incorporating multimodal data significantly enhance CIPN prediction and have the potential for clinical implementation. The use of explainable artificial intelligence techniques further supports their integration into precision oncology. Future research should focus on multicenter validation, real-time EHR integration, and the development of neuroprotective strategies for high-risk patients.