99P Radiologists perception on AI/ML software as a medical device (SaMD) unveiled via post-study usability survey: Key assets to redefine lung cancer screening practice

Abstract

Background: Accurate prognosis in mNSCLC treated with immunotherapy can guide clinical decisions.We developed machine learning (ML) models to predict overall survival (OS) at 6, 12, 18, and 24 months in mNSCLC patients receiving pembrolizumab, based on clinicopathological features. Methods:We retrospectively analysed mNSCLC patients treated with pembrolizumab at the Prof.Dr. Ion Chiricuta Oncology Institute in Cluj-Napoca, Romania, from 2018 to 2023.Patients with missing data or who received <4 cycles of pembrolizumab were excluded, yielding 124 patients.We recorded clinical and paraclinical variables, including age, Eastern Cooperative Oncology Group performance status (ECOG PS), primary tumour T stage, presence of bone or brain metastases, multiple metastatic sites, prior palliative radiotherapy, neutrophil-to-lymphocyte ratio, lactate dehydrogenase, platelet-to-lymphocyte ratio, systemic immuneinflammation index (SII), and PD-L1 status.To address class imbalance, we applied SMOTE.Two SVM models were then trained with an 80/20 train-test split using 5fold cross-validation: a standard SVM and a weighted SVM in which each feature was weighted according to its hazard ratio from a multivariate Cox regression on OS.Results: Model A achieved accuracies of 90%, 82.17%, 82.30%, and 68.50% with areas under the curve (AUC)s of 0.39, 0.84, 0.87, and 0.69 for 6-, 12-, 18-, and 24months OS.Model B reached accuracies of 89.53%, 80.53%, 83.1%, and 69.3% and AUCs of 0.60, 0.86, 0.85, and 0.72 at the same timestamps.SHAP identified ECOG PS, age, multiple metastatic sites, NLR and SII as strongest predictors. Conclusions:These findings indicate that SVM-based algorithms integrating routine clinical data and inflammatory biomarkers can effectively predict survival in mNSCLC patients treated with pembrolizumab.Incorporating Cox-derived prognostic weights for each feature improved model accuracy, suggesting that combining traditional survival analysis with ML enhances discriminative performance.Such AI-driven prognostic tools may help stratify patients by risk and personalise treatment decisions.Validation in larger, independent cohorts is needed to confirm generalisability.

Ähnliche Arbeiten