Interpretable Deep Learning Models for Cancer Diagnosis Using Tumor-Educated Platelet Transcriptomes

Tumor-educated platelets (TEPs) have emerged as promising biomarkers for liquid biopsy-4based cancer diagnostics.However, the intrinsic challenges of transcriptomic data, including high dimensionality, small sample size, and strong nonlinearity, hinder robust classification.In this study, we develop a deep learning model that integrates long short-term memory (LSTM), bidirectional LSTM, convolutional neural network-LSTM (CNN-LSTM), and attention-based LSTM (Attention-LSTM) to classify cancer and healthy samples from the GSE68086 dataset (235 cancer, 50 healthy).Experimental results demonstrate that all models achieve competitive performance, with overall accuracies approaching 0.88, while the Attention-LSTM and CNN-LSTM further enhance interpretability and local feature extraction.Moreover, SHAP analysis confirms that selected genes contributing to classification align with biologically meaningful signals, providing both predictive accuracy and clinical interpretability.These findings highlight the potential of deep sequential models in advancing platelet-based cancer diagnostics.