Comprehensive Summary
Keyl et al. developed an explainable artificial intelligence (xAI) model using multimodal real-world data to improve cancer prognosis prediction. They analyzed 15,726 patients across 38 cancer types, incorporating 350 clinical, imaging, and genetic markers. A deep-learning model trained on this dataset identified 114 key prognostic markers and 1,373 significant interactions. The model was validated on an external cohort of 3,288 non-small cell lung cancer (NSCLC) patients, demonstrating strong reproducibility (r = 0.9, P < 0.001). The xAI model outperformed traditional prognostic tools, achieving an OS concordance index of 0.75 compared to 0.56 for UICC staging (P < 0.001), and showed superior accuracy over ECOG performance status, Charlson Comorbidity Index, and modified Glasgow Prognostic Score. Key predictors of poor outcomes included high C-reactive protein (CRP), particularly when platelet counts were low (Δ RC slope: 0.07, P < 0.001), while factors such as higher free triiodothyronine (fT3), greater abdominal muscle volume, and higher PD-L1 expression were associated with better survival. Kaplan-Meier analysis confirmed the model’s ability to stratify patients into distinct risk groups. The study also found that pan-cancer training improved predictive accuracy compared to single-cancer models, suggesting shared prognostic features across different tumor types.
Outcomes and Implications
This study demonstrates how explainable AI can improve cancer prognosis by considering a wide range of clinical, imaging, and genetic factors rather than relying on traditional staging systems alone. By showing how specific variables (such as CRP levels depending on platelet count or the effects of fT3) contribute to patient outcomes, the model provides a clearer picture of disease progression. The strong validation in lung cancer suggests that similar methods could be applied to other cancer types, helping doctors decide which treatments might be most effective or when more aggressive interventions are needed. The ability to combine data from multiple sources could also help guide treatment decisions in real-time, offering a more complete assessment of each patient’s condition. To be widely used in clinics, the model will need to be integrated into existing medical systems and evaluated for its impact on actual treatment decisions.