Physical Modification of Starch Digestibility: A Comprehensive Review and AI-Assisted Qualitative Knowledge...

Starch modification through physical processing represents a promising "green" strategy to enhance food functionality and nutritional quality while meeting clean-label demands. This review offers a critical overview of the structural and nutritional impacts of physical modifications coupled with an exploratory use of artificial intelligence (AI)-assisted literature for screening and mining. Focused on literature published between 2001 and 2026, a traditional human-led systematic search identifies eligible studies that examined the effects of three major commercially relevant modification techniques-annealing (ANN), heat-moisture treatment (HMT), and autoclaving-on rapidly digestible starch (RDS), slowly digestible starch (SDS), and resistant starch (RS), with emphasis on resistant starch type III (RS3). Among the evaluated techniques, autoclaving, particularly when followed by retrogradation, appears to offer the greatest potential for increasing RS relative to the original native starch. However, its effectiveness is strongly dependent on the raw material, with substantial gains arising from its amylose content, while waxy starches may show little improvement or even a reduction in RS. ANN tends to produce milder and more variable effects, often shifting starch from RDS toward SDS. HMT generally reduces RDS and increases SDS, making it a promising approach for attenuating glycemic responses. However, the extent of these changes can vary substantially depending on the starch source, amylose content, moisture level, and processing temperature. Overall, the work revisits the potential of physical processing as an avenue for starch engineering while underlining a pressing need for standardized workflows, harmonized methodologies, and unified protocols for quantification of starch digestibility, namely, of RS levels. Lastly, the review exemplifies AI can accelerate preliminary literature identification and synthesis yet highlights gaps in AI aptitudes for independent quantitative integration that still maintains a need for thorough human-driven validation.