Towards Reliable AI-Assisted Detection of Sickle Cell Disease: A Machine Learning Framework Accounting for Acquisition Bias, Designed for the Classification of Blood Smear Images

Duplicate views, annotation overlays and acquisition characteristics can bias performance measures, while small microscopy datasets can enable valuable image-classification experiments. This study assesses a reproducible classical machine-learning workflow for the classification of sickle-cell positive and negative blood-smear fields, to include explicit checks for these risks. The provided archive had 991 JPEG files. Positive views were ignored if they were annotated, duplicate pixels were deleted, and groups of candidate duplicate images with conflicting annotations were quarantined. The final analysis consisted of 565 images (421 development; 144 held-out test) split by similarity-group-disjoint partitioning. Five-fold development cross validation was used to compare four image models. We used a separate random forest model designed for prediction of non-biological predictive signal. The hybrid-SVM was chosen prior to the final test scoring, and it obtained accuracy 94.69%, macro-F1 93.19%, sensitivity 95.24%, specificity 93.10% and ROC-AUC 0.961. The accuracy interval for 95% of the grouped-bootstrap values was 90.27–98.33%. The metadata-only model has acquired a ROC-AUC of 0.767, showing acquisition-associated label information. Conclusions: The workflow offers an image-level benchmark reference and an implementation that can be reused. Internal discrimination alone does not provide patient level diagnostic validity and patient-linked and external validation are still required.