Smartphone conjunctival imaging for noninvasive clinical phenotyping and risk stratification of pediatric infectious and hematologic conditions and maternal vascular risk in resource-limited settings

This dissertation examines how far photographs acquired with commercially available smartphones can support noninvasive phenotyping in community and clinical settings and defines the validation required for clinical deployment. The central studies use images of the palpebral or bulbar conjunctiva linked to participant-level reference data. Acquisition, color standardization, segmentation, feature extraction, and model evaluation were designed around participants rather than photographs so that repeated images from one person did not cross development and evaluation partitions.In a cross-sectional cohort of 405 Rwandan children, a conjunctival classifier for malaria rapid diagnostic test positivity achieved an image-level area under the receiver operating characteristic curve (AUC) of 0.76 in a held-out set comprising 1,235 images from 122 participants. A secondary analysis of the publicly released test data (870 images from 81 participants) yielded a participant-level AUC of 0.760 (95% participant-bootstrap CI, 0.629–0.876) after averaging image probabilities within participants. In a separate pediatric anemia study of 565 children, held-out image-level AUCs were 0.77 for palpebral images and 0.79 for bulbar images; the evaluation set contained 170 participants and no participant appeared in both development and evaluation data. These results establish internal discrimination under participant-disjoint evaluation and define etiologic reference testing and prospective clinical utility as the next validation stages.A pregnancy study of 410 participants constructed a CBC-informed image representation comprising 19 positions and 15 unique descriptors, then compared spatial and probabilistic spatiotemporal clustering. The two methods produced corresponding assignments for 364 of 409 classified participants (89.0%), and held-out reassignment accuracy was 77.4% conditional on the fixed representation. Because same-cohort CBC measurements informed feature selection, subsequent hematologic associations are post-hoc biological characterization rather than independent validation; the analysis does not establish iron deficiency, latent anemia, or longitudinal progression.A second pregnancy study analyzed 8,752 images from 413 participants using a prespecified same-visit elevated-blood-pressure threshold of 130/80 mm Hg. In held-out participant evaluation, a radiomics-only neural ensemble achieved an AUC of 0.85, increasing to 0.89 after adding body mass index. In a separate external Visit 1 cohort enrolled in a preeclampsia study, validation of the prespecified analytical method against the participant-specific mean of two cuff readings yielded a held-out participant AUC of 0.833 (95% CI, 0.578–1.000). Repeated-reading concordance analyses retained discrimination across two clinically motivated sensitivity subsets. A parallel analysis in 48 participants found leading palpebral and bulbar radiomic associations of similar absolute magnitude with continuous FMF-estimated preeclampsia risk (|ρ| = 0.451–0.471); no individual feature survived family-wide multiplicity correction.The dissertation also includes a critical review of nonblood malaria detection and a school-based dental-surveillance protocol. The dental study was still enrolling participants and had not produced a completed diagnostic-accuracy analysis. Collectively, the studies establish participant-linked smartphone imaging as a credible clinical measurement approach across infectious, hematologic, and maternal vascular applications. The evidence supports its use for adjunctive risk stratification and quantitative phenotyping within confirmatory care pathways. The dissertation’s principal contribution is the combination of field-realistic acquisition, participant-level analytical control, cross-surface and cross-cohort empirical testing, and an explicit hierarchy linking each result to the clinical claim it can support.