# Artificial intelligence analysis applied to the treatment of granulosa cell tumors of the ovary

## TL;DR

This paper combines mathematical modeling and machine learning to improve predictions of treatment outcomes for rare ovarian tumors called granulosa cell tumors.

## Contribution

A novel hybrid framework integrating mechanistic models with ML improves prediction accuracy for GCT treatment responses.

## Key findings

- Integrating mathematical model outputs improved predictive performance across datasets.
- Neural networks with model-derived variables achieved higher accuracy (up to 91.4%).
- Tumor proliferation and apoptosis rates were the most influential parameters for treatment outcomes.

## Abstract

Granulosa cell tumors (GCTs) of the ovary are rare malignancies with limited systemic treatment options and high recurrence rates. Combining tumor necrosis factor-related apoptosis-inducing ligand (TRAIL)-producing oncolytic viruses with procaspase-3 activator (PAC-1) presents a promising therapeutic strategy, as TRAIL initiates apoptosis while PAC-1 amplifies caspase activity. However, patient responses remain variable, necessitating predictive frameworks that can integrate biological complexity with clinical data.

We developed a hybrid framework that integrates a mechanistic mathematical model of TRAIL-oncolytic virus and PAC-1 therapy with machine learning (ML) algorithms to predict tumor dynamics in GCTs. Four datasets (continuous and categorical tumor size measurements) were analyzed. Clinical and imaging data were merged with individualized solutions from the mathematical model to generate enriched feature sets for ML training. Linear regression and neural network models were trained and evaluated using accuracy, F1 scores, and root mean square error (RMSE).

Integrating mathematical model outputs improved predictive performance across all datasets. Linear regression models showed reduced RMSE compared to models without mathematical features (e.g., RMSE decreased from 18.4 to 16.1 in one dataset). Neural networks incorporating model-derived variables achieved higher accuracy and F1 scores (e.g., accuracy improved from 77.3% to 91.4%). Sensitivity analysis revealed that tumor proliferation and apoptosis rates were the most influential parameters for treatment outcomes.

Our results demonstrate that coupling mathematical modeling with ML enhances the prediction of tumor burden in patients undergoing TRAIL-oncolytic virus and PAC-1 therapy. This integrative approach provides mechanistic insight into tumor behavior while improving predictive accuracy, supporting the development of personalized therapeutic strategies for GCTs. The framework also offers broader applicability to other cancers with limited treatment options and heterogeneous responses.