Few-Shot, Transfer, and Meta-Learning in Naturally Data-Scarce Vertical Domains: A Cross-Sector Methodological Survey from Rare-Disease Diagnosis, Emerging-Financial-Product Risk Management, Autism Spectrum Variability, and Low-Resource Language Processin

Abstract

Vertical applications of artificial intelligence increasingly run into a structural barrier that does not yield to better engineering alone: the underlying population is small, heterogeneous, or simply not well represented in any historical archive. Rare-disease cohorts assemble dozens to a few hundred patients across continents; freshly issued credit instruments such as buy-now-pay-later receivables, climate-linked structured notes, and digital-asset derivatives have at most a few quarters of post-origination behaviour; behavioural signals collected from children on the autism spectrum vary so much across individuals that a model trained on one child often fails on the next; and several hundred living languages have no standard parallel corpus large enough to train conventional encoders. We refer to these settings collectively as naturally data-scarce verticals, and we ask whether the few-shot, transfer, and meta-learning literature accumulated across otherwise unrelated fields exposes a transferable methodological pattern. To answer this we curated 142 recent studies that span clinical decision support, financial-network monitoring, autism-skill acquisition, multilingual healthcare communication, privacy-preserving multi-institution learning, multimodal fusion, and explainable risk attribution. We extract a common backbone — pretrain → adapt → calibrate → audit — and operationalise it through a unified evaluation protocol that combines four simulated benchmarks. Across 36 model–scenario combinations, parameter-efficient adaptation with prototype heads achieves the best accuracy–stability trade-off when per-class support is below 20, while metric-based meta-learning dominates when classes are rare but per-class support is moderate. Calibration matters as much as raw accuracy: post-hoc temperature scaling combined with Bayesian last-layer recalibration reduced the expected calibration error from 0.187 to 0.046 on the financial-instrument benchmark. We close with a discussion of fairness, privacy, and deployment audit obligations and sketch a research agenda built around shared benchmarks, federated meta-learning, and lifecycle-aware reliability accounting.