Inferring Individualized Color-Vision Distortions from fMRI Hue-Representation Geometry

Many health-related data carry structured distortions that can act as phenotypes, yet interpreting such signatures and translating them into intervention remains challenging. Color vision deficiency (CVD), for example, distorts pairwise relations between neural hue representations while categorical recognition remains relatively preserved. Here we formulate fMRI-measured hue responses as structured neural representations and infer low-dimensional distortions that explain their geometry. We quantify this distortion using the Leave-One-Color-Out (LOCO) interpolation vulnerability profile, which measures how well each held-out hue can be interpolated from the remaining hue manifold. Applied to Ishihara-confirmed CVD participants across visual areas (V1-hV4), the framework recovers subject-specific distortions through both a one-parameter retinal cone-shift model and a two-parameter retinal–cortical model. Finally, we show that retinal and cortical models converge on detection performance but prescribe divergent corrections, yielding concrete predictions for planned behavioral validation. These findings show that detection-level agreement is insufficient for choosing an intervention and demonstrate that structured distortions in high-dimensional biological data can be resolved into low-dimensional, subject-specific models for personalized intervention.