Treatment-Conditioned Diffusion for Forecasting Neurodegenerative Disease Progression

Danylo Boiko, Viktoriia Mishkurova · 2026

Read Treatment-Conditioned Diffusion for Forecasting Neurodegenerative Disease Progression by Danylo Boiko, Viktoriia Mishkurova (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-29932.txt.

Introduction

Danylo Boiko and Viktoriia Mishkurova forecast brain disease advance by picturing future brain states, not single scores. Their treatment-aware diffusion model starts from screening DaTscan and a year of levodopa dose, tracking drug effects over time with a Transformer encoder. A weighted mask keeps focus on brain zones that matter for the disease. The model holds sharp anatomical edges and beats the baseline, cutting mean squared error about fourteen percent and lifting image likeness about five percent. The result points toward care plans tuned to each patient's course.

Catalog record

CREATORDanylo Boiko, Viktoriia Mishkurova
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE29169
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.29932

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Treatment-Conditioned Diffusion for Forecasting Neurodegenerative Disease Progression Danylo Boiko, Viktoriia Mishkurova Abstract Forecasting the progression of neurodegenerative diseases, such as Parkinson’s disease, is essential for effective long-term planning and personalized therapeutic intervention. Existing systems typically produce scalar clinical scores that ignore the rich structure of longitudinal neuroimaging, while traditional generative approaches suffer from a loss of anatomical details and blurring subtle progression patterns. To address this, we introduce a novel treatment-conditioned diffusion framework that predicts high-fidelity future brain states by conditioning the generative process on patients’ screening DaTscan images and levodopa equivalent daily dose over one year. The pipeline uses a Transformer-based encoder to represent non-linear, time-dependent pharmacological dynamics and optimizes generation through a multi-weight region-of-interest mask that focuses on biologically critical areas. Experimental evaluation shows that our framework maintains sharp anatomical boundaries and significantly improves clinical fidelity relative to the baseline, achieving 14.0% lower MSE, 7.2% lower MAE, and 4.9% higher SSIM. 1 Introduction Modeling the progression of neurodegenerative diseases, including Parkinson’s, is a key factor in the personalization and long-term planning of therapeutic strategies. Parkinson’s disease (PD) is one of the leading neurodegenerative diseases worldwide and the fastest-growing among them 8 . This condition is associated with a wide range of motor and non-motor symptoms that significantly reduce quality of life. Early diagnosis and therapeutic intervention are correlated with slower progression and delayed onset of certain symptoms. Classical predictive techniques ranging from multilayer perceptrons (MLPs) and recurrent neural networks (RNNs) to more recent graph-based or ordinary differential equation (ODE) approaches typically output scalar clinical scores ( e.g . , UPDRS, MMSE, MoCA) while ignoring the rich structure of longitudinal neuroimaging ( e.g . , MRI, CT, PET, SPECT). Generative approaches such as generative adversarial networks (GANs) and variational autoencoders (VAEs) are useful for synthesizing brain images 31 . However, they remain plagued by loss of anatomical details and fuzziness of subtle progression patterns. In this work, we introduce a novel treatment-conditioned diffusion paradigm. Building on the success of conditional diffusion models (CDMs), which have achieved state-of-the-art results in high-fidelity medical image synthesis 36 , we propose leveraging their generative power to forecast PD progression by conditioning the diffusion process on patients’ data. In particular, the model learns the conditional distribution based on DaTscan images, a type of SPECT imaging used to evaluate dopamine transporters, and the monthly levodopa equivalent daily dose (LEDD) over the course of a year. To implement this framework, we designed a lightweight 2D U-Net 29 that predicts future DaTscan images by channel-wise concatenation of noise and screening DaTscan images, while incorporating LEDD embeddings via a projection layer. Key contributions are highlighted below: • Novel diffusion framework that forecasts brain states conditioned on screening DaTscan images and treatments. • Contrastive Transformer-based autoencoder for noisy, time-dependent medication histories. • Multi-weight ROI optimization pipeline that forces the model to prioritize biologically critical regions. • Improvement of clinical fidelity, achieving 14.0% lower MSE, 7.2% lower MAE, and 4.9% higher SSIM across 14 slices compared to the baseline. 2 Related work Over the last decades, the mathematical and computational modeling of neurodegenerative disease progression has shifted from linear clinical observations toward complex longitudinal frameworks that account for multi-modal biological decay. The central element of this evolution is the merging of diverse data streams ranging from longitudinal imaging and genetic markers to cerebrospinal fluid (CSF) proteomics 15 ; 28 . These models assume that disease-specific biomarkers follow predictable paths that can be captured through latent-variable modeling or state-space representations 30 ; 23 . By synchronizing patients at different clinical stages, we can identify early-stage physiological changes that precede a clear symptomatic decline, providing a critical window for medical treatment 33 ; 23 . Although statistical methods provide interpretability, recent advances in machine learning (ML) and deep learning (DL) offer a notable capacity to handle the high-dimensional, non-Gaussian nature of clinical datasets 28 . For example, graph neural networks (GNNs) and neural ODE models are increasingly utilized to predict individual-level brain atrophy or cognitive scores by capturing long-range dependencies in longitudinal data 14 . These approaches excel in identifying subtle, non-linear relationships between risk factors, such as the combined effect of genetic mutations ( e.g . , in GBA or LRRK2) and motor symptom progression 15 ; 33 . However, a persistent limitation of these models is their lack of a biological basis, unlike mechanistic biophysical models 30 . Consequently, there is a growing convergence toward hybrid systems that constrain ML outputs with known physiological priors to ensure clinical relevance 14 ; 10 . GANs and VAEs demonstrated excellent capabilities for cross-modality synthesis and the augmentation of small datasets, enabling the creation of high-accuracy anatomical transitions 9 . This predictive power is further refined by denoising diffusion probabilistic models (DDPMs), which provide a mathematically solid framework for generating diverse synthetic brain images while maintaining the structural integrity of pathological markers 11 . Despite the strength of these models in capturing non-linear morphological changes, a major area of convergence remains the requirement to balance computational complexity with clinical interpretability, ensuring that synthetic outputs align with the biological priors of neurodegeneration. Nowadays, applications of diffusion models have evolved from image synthesis to forecasting the progression of neurodegenerative diseases, offering a robust probabilistic framework to predict future brain states 37 . While a substantial number of systems used this technique for cross-sectional normative modeling 34 and unsupervised anomaly detection 3 ; 4 , the methodological focus is rapidly expanding toward longitudinal predictive modeling. To enforce precise anatomical constraints throughout image generation, novel pipelines use segmentation-guided sampling and mask-aware training, ensuring high fidelity while allowing localized flexibility 12 . By framing disease progression as a conditional generation task, CDMs can synthesize highly realistic future neuroimaging representations that reflect anticipated pathological changes, overcoming the weaknesses of deterministic predictive algorithms 6 ; 35 . The main methodological assumption here is that patients’ data provide a sufficient predictive signal to dictate localized, time-dependent pathological deformations. Consequently, while generative models show significant strengths in forecasting subject-specific trajectories with narrow confidence intervals, their predictive accuracy can degrade while forecasting highly non-linear neurodegeneration or unobserved variants 35 . 3 Data and preprocessing Data used in this paper were obtained on September 28, 2025, from the Parkinson’s progression markers initiative (PPMI) database 18 , an ongoing longitudinal observational study launched in 2010 by the Michael J. Fox Foundation for Parkinson’s research. PPMI aims to identify biomarkers for the onset and progression of PD by collecting multimodal data from diverse cohorts. Levodopa is the most effective and widely prescribed treatment for PD. It is a precursor to dopamine that can pass through the blood-brain barrier, thus restoring dopamine levels in the striatum, reducing motor symptoms, and improving patients’ quality of life 21 . We selected LEDD as a condition because it reflects the cumulative dopaminergic effect on the brain throughout the day, rather than temporary peaks from single doses. Using DaTscan images in combination with LEDD allows to directly correlate structural and functional changes in the basal ganglia with the patient’s clinical condition. Moreover, by focusing on slices 34 to 47, which cover the caudate nucleus and putamen, we target the most affected areas while reducing noise from less informative regions, increasing the

Read the complete text