Malaria microscopy becomes clinically risky when parasitized erythrocytes are missed, even when a classifier reports high overall accuracy. This study presents an artificial-intelligence-assisted malaria microscopy framework that explicitly optimizes the diagnostic operating point for sensitivity-aware screening. A balanced set of 27,558 National Institutes of Health/National Library of Medicine (NIH/NLM) thin-smear cell images was divided by stratified image-level sampling in the ratio 70:15:15 for training, validation, and independent testing. Two custom convolutional neural networks (CNNs) with complementary capacity–regularization profiles and an equal-weight score-level ensemble were evaluated. Training-only online augmentation, fixed input resizing, validation-only threshold selection, confidence intervals, and error analysis were incorporated to improve reproducibility. At the conventional threshold, the compact CNN achieved the highest accuracy (95.26%). After optimizing the operating point with the recall-weighted F2 objective, the deeper CNN provided the strongest screening-oriented result on the held-out test set, reaching 97.05% sensitivity (95% confidence interval: 96.23%–97.70%), a 96.01% F2-score, a 2.95% false-negative rate, and a receiver-operating-characteristic area under the curve of 0.9876 (95% confidence interval: 0.9842–0.9910). It reduced missed parasitized cells from 177 to 61 relative to its default threshold. The public release does not provide patient identifiers; therefore, the reported split is image-level rather than confirmed to be patient-level independent, and external multicenter validation remains necessary before clinical deployment. The main finding is that the highest-accuracy classifier is not necessarily the most appropriate sensitivity-oriented screening configuration.