Transforming fnirs analysis through AI-assisted simulation, calibration, and interpretation

The National Institute of Biomedical Imaging and Bioengineering awarded Rensselaer Polytechnic Institute $366,006 on September 1, 2026, under the Discovery and Applied Research for Technological Innovations to Improve Human Health program (CFDA 93.286) to develop a simulation-based inference framework that transforms functional near-infrared spectroscopy (fNIRS) data analysis through artificial intelligence.

The project delivers a biophysically detailed simulator integrating realistic models of hemodynamic response, systemic physiology, and data corruption using mesh-based Monte Carlo methods. The framework calibrates this simulator to individual trials and subjects using amortized Bayesian inference to enable fast, accurate estimation of latent neural parameters with uncertainty quantification across experimental and hardware setups. The work generates large-scale synthetic datasets that preserve subject- and task-specific dynamics to support data augmentation for deep learning and interpretable downstream analysis. Semi-supervised learning tools merge synthetic and real data, diagnose simulation gaps, and improve reliability under distribution shifts. Hierarchical Bayesian models extract trial- and subject-level parameters for noise-resilient subject-specific brain understanding. Computational benchmarks and model applications address surgical training scenarios using raw and denoised time series and biophysical parameters for interpretable, uncertainty-aware skill classification and performance prediction.

The award addresses persistent methodological challenges in fNIRS—including sensitivity to physiological noise, motion artifacts, individual variability, and lack of standardized, scalable analysis pipelines—that currently limit broader adoption in research and clinical settings.

Performance occurs in Troy, New York, with a completion date of June 30, 2030.

Abstract

Functional near-infrared spectroscopy (fnirs) has emerged as a powerful tool for monitoring brain activity in naturalistic environments. However, its broader adoption in research and clinical settings is limited by persis- tent methodological challenges that include sensitivity to physiological noise, motion artifacts, individual vari- ability, and a lack of standardized, scalable analysis pipelines and models. This project aims to fundamentally transform how fnirs data is modeled, interpreted, and used by developing the first comprehensive simula- tion-based inference framework tailored to fnirs. At the core of the proposed approach is a biophysically detailed simulator that integrates realistic models of the hemodynamic response, systemic physiology, and data corruption based on mesh-based monte carlo (MMC). We will calibrate this simulator to individual trials and individuals using amortized bayesian inference (ABI), enabling fast, accurate estimation of latent neural parameters with full uncertainty quantification across experimental and hardware setups. Leveraging these calibrated models, we will generate large-scale SYN- thetic datasets that preserve subject- and task-specific dynamics, supporting robust data augmentation for deep learning and interpretable downstream analysis. Our framework also introduces semi-supervised learning tools to merge synthetic and real data, diagnose simulation gaps, and improve reliability under distribution shifts. Further, we will implement hierarchical bayesian models to extract trial- and subject-level parameters for noise-resilient subject-specific brain under- standing. We will provide several computational benchmarks and model skill acquisition in surgical training scenarios, using both raw and denoised time series and biophysical parameters for interpretable, uncertainty- aware skill classification and performance prediction. This project will yield an open-source software suite that facilitates real-time model calibration, data augmen- tation, interpretable AI, and integrative modeling for fnirs. By addressing key limitations in current fnirs workflows, our work will significantly advance the state of the art in computational neuroimaging and support broader applications in clinical neuroscience, neuroergonomics, and personalized brain-computer interfaces.