2P - Using machine learning (ML) and explainable artificial intelligence (AI) to accurately predict immune-checkpoint inhibitor (ICI) response in small cell (SCLC) and non-small cell (NSCLC) lung cancer patients

Background

ICIs are widely used in SCLC and NSCLC, however current predictors of response, such as PD-L1 expression, have limited accuracy. Until recently, identifying predictors for ICI response has been challenging. Utilising ML algorithms, we have developed models to predict ICI outcomes that can facilitate precision medicine and accelerate drug development.

Methods

471 datasets were included for patients commencing ICIs between 2015-2024 in a U.K hospital. Chi-squared analysis was used for comparative statistics. Six ML algorithms were developed using 80% of the data for training, and 20% for validation. Shapley Additive Explanations (SHAP) explainable AI was used for model interpretation.

Results

Of the 471 patients, 431 (92%) were treated with palliative intent and 40 (8%) with curative intent. 430 (91%) had NSCLC and 41 (9%) had SCLC. Patients who had ICI after previous treatment with radiotherapy and chemotherapy had a significantly higher likelihood of progressive disease (PD) (37%) compared to patients who had ICI alone without prior radiotherapy and chemotherapy (18%) (x²=9, p=0.003). XGBoost classifier predicted PD with 78% accuracy. Important features included increasing age, advanced TNM staging, performance status and female sex. The largest patient cohort in our dataset were 390 NSCLC patients who received treatment with palliative intent. In this cohort, the commonest ICI was pembrolizumab (302, 77%) followed by atezolizumab (62, 16%). A sub-group analysis of this cohort yielded an accuracy of 71% and found that prior radiotherapy and atezolizumab use were important features in predicting PD. Our ML algorithm predicted ICI outcomes with significantly greater accuracy (78%) than PD-L1 levels alone (21%) (x²=298, p=0.001).

Conclusions

We developed ML algorithms that accurately predicted ICI response in patients with lung cancer. This could enable informed, shared decision-making between clinicians and patients considering initiating, continuing or stopping ICI therapies. Future work will validate the AI model by incorporating it into ICI drug development trials to prospectively predict patients' responses.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Curenetics.

Disclosure

All authors have declared no conflicts of interest.