Optical Diffraction Tomography and Interpretable Machine Learning Reveal Biophysical Signatures of Gametocyte-Stage Malaria in Red Blood Cells

The study presents and validates an ODT-plus-interpretable-machine-learning framework for label-free, stage-resolved characterization and classification of malaria-infected red blood cells. Relative to conventional tools relying on Giemsa-stained microscopy and rapid diagnostic tests (RDTs), this work acquires three-dimensional refractive index tomograms in a label-free manner and simultaneously outputs quantifiable morphological and biophysical descriptors. Based on three-dimensional refractive index tomograms of RBCs from synchronized P. falciparum cultures, extracting physically interpretable descriptors including sphericity, solidity, eccentricity, dry mass, and maximum refractive index, and integrating self-supervised ViT image representations for downstream analysis.

Gametocyte-stage infected RBCs exhibit distinct morphology, with significantly reduced sphericity and increased eccentricity compared to ring-stage infected or normal RBCs. This finding quantifies stage-specific morphological differences into measurable biophysical metrics, whereas conventional methods struggle to quantify or differentiate infection stages in real time. Derived from ODT three-dimensional refractive index tomograms of synchronized cultured RBCs, with group comparisons using extracted morphological descriptors; the abstract reports significant differences.

Supervised models trained on the combined feature space achieve 88.3% accuracy in multiclass classification and 98.1% in binary classification. This result demonstrates the feasibility of integrating physically interpretable descriptors with data-driven image representations for stage-resolved classification, whereas existing diagnostic tools are limited in real-time stage differentiation. Supervised models trained on the combined feature space, reporting 88.3% accuracy for multiclass (normal, ring, gametocyte) and 98.1% for binary (normal vs infected) classification.

SHAP analysis reveals deep-learning-derived features as the strongest predictors for normal and ring-stage RBCs, while handcrafted features critically influence gametocyte-stage classification. This interpretability analysis links classification decisions to pronounced morphological changes in gametocyte-stage RBCs, providing transparent insight into the basis of model predictions. SHAP analysis attributing predictor importance within the combined feature space, with the abstract reporting differences in feature contributions across stages.

This work targets RBCs from synchronized P. falciparum cultures, achieving label-free stage-resolved characterization and classification under controlled experimental settings, applicable to distinguishing ring and gametocyte stages; potential application settings include high-throughput analysis, culture monitoring, and future diagnostic workflows, though the abstract does not state whether it has been validated on clinical samples or across different laboratory conditions.

Readers may watch: the generalizability of the combined feature space to independent datasets or clinical samples; whether the relative contributions of handcrafted and ViT features across stages remain stable; the compatibility of ODT throughput with culture monitoring scenarios; and whether SHAP attributions remain consistent across different models or data splits. The abstract does not provide sample sizes, confidence intervals, or external validation results, which are open questions for assessing translational potential.