CHIASM: A Self-Supervised Visual Field Encoder for Neuro-Ophthalmology

Background: Artificial intelligence (AI) systems for glaucoma diagnosis and prognostication from visual fields (VF) are under active development, yet do not audit for vertical-meridian-respecting field loss - known sequelae of stroke, hemorrhage, and neoplasm. We developed a self-supervised encoder of automated perimetry that learns anatomically interpretable VF structure without labels, and evaluated its capacity to identify suspected neurologic VF patterns in an independent public glaucoma dataset. Methods: We pretrained a 128-dimensional masked autoencoder on 23,223 unlabeled Humphrey VFs (patient-grouped training split of 28,943 fields from 3,871 patients; UWHVF, all-comers perimetry), using monocular pattern-deviation input. A supervised linear classifier over vertical-midline latent dimensions was trained on per-eye expert neurological/non-neurological labels and assessed under hard-negative cross-validation, with specificity evaluated on 100 held-out, structurally separated UWHVF controls. External evaluation used the Harvard-Glaucoma Fairness dataset (Harvard-GF; 3,300 patients with paired VF and optical coherence tomography [OCT] from a single academic center), which contributed no data at any training stage. Results: Masked reconstruction recovered structure concordant with retinal neuroanatomy: 50 of 128 latent dimensions emerged spatially specialized, versus 23 for the total-deviation encoder. The classifier achieved cross-validated balanced accuracy 0.78 (95% CI, 0.75-0.82) and AUC 0.85 (95% CI, 0.82-0.89), with no false positives among the 100 held-out controls. Applied to Harvard-GF without fine-tuning, it identified a top-20 of 1,748 glaucoma-labeled patients (1.1%) with morphology inconsistent with glaucoma; all 20 were positive on the rule-based Neurological Hemifield Test (mean score 62.4), and OCT showed preserved superior (Cohen d = +0.68; P < .001) and inferior (d = +0.63; P = .003) retinal nerve fiber layer versus severity-matched controls. Conclusions: A self-supervised VF encoder learned anatomically interpretable visual field structure from unlabeled data and identified suspected neurological cases in a curated glaucoma dataset, with expert, rule-based, and OCT corroboration. Visual field datasets used to train glaucoma AI may benefit from neurological screening before model training; the encoder reported here supports such audits and provides a foundation for neuro-ophthalmic AI beyond fundus photography and OCT.