[Paper Review] A Patient-Specific Digital Twin for Adaptive Radiotherapy of Non-Small Cell Lung Cancer

A. Sud, Jialu Huang+5 more|arXiv (Cornell University)|Feb 15, 2026

Radiomics and Machine Learning in Medical Imaging0 citations

TL;DR

The paper develops COMPASS, a patient-specific digital twin for adaptive radiotherapy in NSCLC, using per-fraction multimodal data to predict toxicity and guide treatment with a GRU autoencoder and logistic regression.

ABSTRACT

Radiotherapy continues to become more precise and data dense, with current treatment regimens generating high frequency imaging and dosimetry streams ideally suited for AI driven temporal modeling to characterize how normal tissues evolve with time. Each fraction in biologically guided radiotherapy(BGRT) treated non small cell lung cancer (NSCLC) patients records new metabolic, anatomical, and dose information. However, clinical decision making is largely informed by static, population based NTCP models which overlook the dynamic, unique biological trajectories encoded in sequential data. We developed COMPASS (Comprehensive Personalized Assessment System) for safe radiotherapy, functioning as a temporal digital twin architecture utilizing per fraction PET, CT, dosiomics, radiomics, and cumulative biologically equivalent dose (BED) kinetics to model normal tissue biology as a dynamic time series process. A GRU autoencoder was employed to learn organ specific latent trajectories, which were classified via logistic regression to predict eventual CTCAE grade 1 or higher toxicity. Eight NSCLC patients undergoing BGRT contributed to the 99 organ fraction observations covering 24 organ trajectories (spinal cord, heart, and esophagus). Despite the small cohort, intensive temporal phenotyping allowed for comprehensive analysis of individual dose response dynamics. Our findings revealed a viable AI driven early warning window, as increasing risk ratings occurred from several fractions before clinical toxicity. The dense BED driven representation revealed biologically relevant spatial dose texture characteristics that occur before toxicity and are averaged out with traditional volume based dosimetry. COMPASS establishes a proof of concept for AI enabled adaptive radiotherapy, where treatment is guided by a continually updated digital twin that tracks each patients evolving biological response.

Motivation & Objective

• Motivate the need for dynamic, patient-specific modeling of normal tissue toxicity in NSCLC treated with biologically guided radiotherapy (BGRT).
• Develop COMPASS (Comprehensive Personalized Assessment System) as a temporal digital twin to integrate per-fraction imaging, dosimetry, and dose–response kinetics.
• Demonstrate a learning pipeline that predicts CTCAE grade 1+ toxicity using organ-specific latent trajectories learned by a GRU autoencoder and classified by logistic regression.
• Provide evidence that AI-driven early toxicity warnings can emerge several fractions before clinical toxicity.
• Show how dense BED-driven representations capture spatial dose texture characteristics related to toxicity that traditional dosimetry may miss.

Proposed method

• Aggregate per-fraction PET, CT, dosiomics, radiomics, and cumulative BED kinetics for each patient as time-series data.
• Train a GRU autoencoder to learn organ-specific latent trajectories from the temporal data.
• Classify latent trajectories with logistic regression to predict CTCAE grade 1 or higher toxicity.
• Analyze eight NSCLC patients with BGRT to obtain 99 organ-fraction observations across 24 organ trajectories (spinal cord, heart, esophagus).
• Assess the timing and characteristics of toxicity risk signals and compare BED-driven representations to traditional dose-volume approaches.

Experimental results

Research questions

• RQ1 Can a patient-specific temporal model predict NSCLC normal tissue toxicity earlier than conventional methods?
• RQ2 Do organ-specific latent trajectories derived from multimodal, time-series data provide meaningful toxicity signals for adaptive radiotherapy?
• RQ3 Is BED-driven texture information associated with impending toxicity beyond conventional dosimetric metrics?
• RQ4 What is the feasibility of a digital-twin approach (COMPASS) to continuously update treatment decisions in BGRT?
• RQ5 How does the approach perform in a small but temporally rich NSCLC cohort?

Key findings

• The COMPASS framework supports AI-driven early warning of toxicity several fractions before clinical signs appear.
• Dense BED-driven representations reveal spatial dose texture characteristics linked to toxicity that may be averaged out by traditional dosimetry.
• GRU autoencoder learned organ-specific latent trajectories that, when classified, predict CTCAE grade 1+ toxicity.
• The study analyzes eight NSCLC patients with 99 organ-fragment observations across 24 organ trajectories, demonstrating the feasibility of dense temporal phenotyping for adaptive radiotherapy.
• The results establish proof of concept for a continually updated digital twin guiding personalized radiotherapy decisions.