Abstract 2467: Advancing multimodal machine learning techniques to predict platinum resistance in high-grade serous ovarian cancer

Abstract Background: Median survival in high-grade serous ovarian cancer (HGSOC) drops to <2 years when it advances rapidly and is declared platinum-resistant (PR). Prior efforts using machine learning to predict PR from hematoxylin and eosin (H&E) whole slide images (WSIs) have shown modest success with an area under the curve (AUC) ∼0.6. Current Computational Pathology (CPath) methods frequently overfit due to relative data scarcity and dimensionality mismatches between gigapixel-resolution WSIs and a binary outcome. This project aims to improve CPath performance in predicting PR in newly diagnosed HGSOC by optimizing multimodal fusion of clinical data elements. Methods: We defined PR as progression or recurrence within 12 months of completing platinum-based therapy, and platinum sensitivity (PS) as recurrence > 12 months after completing treatment. The model was trained on 243 H&E WSIs from 127 patients (62 PR, 65 PS) with repeated 5-fold cross-validation and validated on The Cancer Genome Atlas-Ovary (TCGA) dataset comprising 1156 WSIs from 578 patients (188 PR, 390 PS). H&E WSIs at 20x magnification were tiled, with features extracted via the UNI foundation model and analyzed using attention-based multiple-instance learning to predict PR. FIGO stage and largest residual at primary cytoreduction were used as clinical parameters and fused into the model using different strategies - early, intermediate via concatenation, intermediate via Kronecker product (KP), and late - and performance compared across techniques. Results: On the internal test set, intermediate-fusion via KP achieved the highest AUC and balanced accuracy (bAcc) (AUC 0.806, bAcc 0.744), followed by early (AUC 0.726, bAcc 0.661), intermediate via concatenation (AUC 0.680, bAcc 0.617), late (AUC 0.672, bAcc 0.628) and the model without clinical data (AUC 0.629, bAcc 0.593). On the TCGA dataset, the best-performing model had an AUC of 0.622 and bACC of 0.566, compared to published baseline AUC of 0.602 (Ahn et al, May 2024) which was trained on 754 WSIs of 394 patients (132 PR, 262 PS). Conclusions: Despite a smaller training set, we surpass recently published models for predicting PR in HGSOC from H&E WSIs by adding simple clinical data points, favoring an interwoven intermediate fusion technique. Multimodal fusion via KP thus represents a means to expand the use of CPath to predict clinical outcomes in diseases with relative data scarcity. These results suggest progress towards better matching cases of PR HGSOC to non-platinum treatment. Citation Format: Leonhard Donle, Kelsey C. Goon, Ernst Lengyel. Advancing multimodal machine learning techniques to predict platinum resistance in high-grade serous ovarian cancer [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 2467.