AI-Powered TwinWeaver Framework Improves Clinical Event Forecasting in Oncology

Global: AI-Powered TwinWeaver Framework Improves Clinical Event Forecasting in Oncology

A team of researchers has introduced TwinWeaver, an open-source framework that converts longitudinal patient records into textual format, enabling large language models to predict clinical events and forecast trajectories across diverse cancer types. The approach is detailed in a preprint posted on arXiv, where the authors report testing the system on 93,054 patients representing 20 different cancers.

Framework Overview

TwinWeaver serializes multi-modal, sparse clinical time series into a unified text representation, allowing a single model to handle event prediction, survival analysis, and therapy‑switch forecasting without separate specialized pipelines.

Benchmark Performance

In comparative benchmarks, the TwinWeaver‑based Genie Digital Twin (GDT) achieved a median Mean Absolute Scaled Error (MASE) of 0.87, outperforming the strongest traditional time‑series baseline, which recorded a median MASE of 0.97 (p<0.001). The improvement reflects more accurate short‑ and long‑term forecasting across the evaluated cohort.

Risk Stratification Gains

Beyond raw forecasting error, GDT enhanced risk stratification, attaining an average concordance index (C‑index) of 0.703 across survival, disease progression, and therapy‑switching tasks. This surpasses the best baseline model, which achieved a C‑index of 0.662, indicating better discrimination of high‑risk patients.

Generalization to Clinical Trials

The authors further tested GDT on out‑of‑distribution clinical trial data. In zero‑shot settings, GDT matched baseline performance, while fine‑tuning yielded median MASE values ranging from 0.75 to 0.88 and an average C‑index of 0.672, exceeding the strongest baseline’s C‑index of 0.648.

Interpretability and Clinical Reasoning

TwinWeaver also supports an interpretable clinical‑reasoning extension, offering transparent insights into model predictions and facilitating scalable deployment in longitudinal clinical modeling scenarios.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.