# MMHC-OCPR: Prediction of Platinum Response and Recurrence Risk in Ovarian Cancer with Multimodal Deep Learning

## TL;DR

This paper introduces a deep learning model that predicts platinum response and recurrence risk in ovarian cancer patients using pathology images and clinical data, aiming to improve personalized treatment.

## Contribution

A novel multimodal deep learning model (MMHC-OCPR) is developed for predicting platinum resistance and recurrence risk in ovarian cancer.

## Key findings

- The model achieved an AUC of 0.914 for platinum response prediction when integrating metastatic images and clinical data.
- Recurrence risk prediction reached a C-index of 0.838 with multimodal input.
- Patients were stratified into three risk groups with distinct 2-year progression-free survival rates.

## Abstract

Background/Objectives: Ovarian cancer has the highest mortality among gynecological malignancies, with platinum resistance significantly contributing to poor prognosis. We aimed to develop a multimodal model (MMHC-OCPR) to predict platinum response and recurrence risk, enabling earlier personalized treatment and improved outcomes. Methods: This multicenter retrospective study included a combined cohort of 431 patients, comprising 1182 whole slide images (WSIs) curated from two independent datasets. The primary cohort consisted of 376 patients from the National Cancer Center (China), which was further partitioned into training, validation and internal test sets to ensure model development and evaluation. An additional external test cohort was incorporated using publicly available data from TCGA, enhancing the generalizability of our findings. We implemented a weakly supervised multiple instance learning framework to integrate histopathological imaging with clinicopathological variables, further strengthened by the incorporation of the transformer-based pretrained encoder UNI2-h, which enhanced the model’s predictive performance. Results: All patients in the primary cohort had pathology slides collected from primary ovarian tumors and metastatic tumor, along with clinical factors related to prognosis and treatment response. The baseline platinum response classifier using primary WSIs achieved an AUC of 0.896 in the internal test group and 0.876 in the external test group. Integration of metastatic WSIs and clinical data inputs yielded a superior AUC of 0.914 in the internal test set. The recurrence risk model demonstrated a C-index of 0.801, rising to 0.838 after multimodal enhancement. The model stratified patients into low-, intermediate- and high-risk groups with 2-year progression-free survival rates of 77.3%, 48.0% and 2.0%, respectively. Conclusions: Our model enables the early detection of platinum resistance, guiding timely treatment intensification. The recurrence risk stratification supports personalized management by identifying patients with favorable outcomes following surgery and chemotherapy, potentially sparing them from maintenance therapy to reduce associated toxicity, cost, and enhance quality of life.