HPExplainPro: A Framework for Pan-cancer Prognosis Prediction Based on Deep Interpretable Learning
Keywords: Interpretable Learning, Deep Learning, Black Box, Prognosis Prediction, Pan-cancer
TL;DR: HPExplainPro: A Framework for Pan-cancer Prognosis Prediction
Abstract: Prognostic prediction research holds immense significance in guiding doctors in evaluating the effectiveness of various treatment modalities, thereby facilitating the selection of the most appropriate treatment plan tailored to individual patients. However, a critical challenge that persists in this domain is the scarcity of clinical interpretability. This issue primarily stems from the inherent opacity of deep learning algorithms, often referred to as "black boxes," where most models operate with closed decision-making processes, lacking transparency in explaining the reasons behind their predictions. To address this gap, we designed HPExplainPro, a deep explainable learning framework for pan-cancer prognostic prediction. HPExplainPro is composed of a deep learning model rooted in expert knowledge, a data-driven feature fusion approach, a triple feature selection technique, a heterogeneous classifier, and a secondary learning probability error integration model. At its core, HPExplainPro features the Deepxplain module, which leverages global interpretation via DeepSHAP and local interpretation through LIME algorithms to provide insights into the decision-making process. To demonstrate the superiority of HPExplainPro, this article employs three distinct cancer datasets sourced from preeminent hospitals in China. These datasets were leveraged to construct an immunotherapy ORR prediction model for lung cancer, a 5-year survival prediction model for breast cancer patients, and a local progression outcome prediction model for early liver cancer microwave ablation. The experimental results unequivocally demonstrate that HPExplainPro outperforms alternative methods. Furthermore, through Deepxplain's global interpretation capabilities, the study identifies potential prognostic biomarkers such as NETs, LDH, and NLR, which significantly influence the outcome of lung cancer immunotherapy. Additionally, HPExplainPro's local interpretation functionality enables individualized prognostic predictions for lung cancer patients, offering clinicians tailored insights into patient-specific responses to treatment, see in figure 1. Beyond lung cancer, this article explores the broader applicability of HPExplainPro in other diseases. Specifically, it presents a COVID-19 critical illness prediction model using patient data from Wuhan Third Hospital, illustrating the flexibility of HPExplainPro in addressing diverse clinical challenges. Additionally, the study delves into the utilization of HPExplainPro in prognostic prediction within the realm of traditional Chinese medicine (TCM), developing a prognostic prediction model, named "Zhongjing," that harmoniously integrates principles from both TCM and Western medicine. This additional validation further underscores the generalization performance of HPExplainPro across different medical domains. Lastly, the reliability and practicality of HPExplainPro are further bolstered by its validation in the breast department of Hunan Provincial Cancer Hospital. This comprehensive validation process not only validates the effectiveness of HPExplainPro but also showcases its potential to enhance clinical decision-making and improve patient outcomes across a wide range of cancers and diseases.
Submission Number: 22
Loading