Performance diagnostique, gain de temps réel et utilité clinique d’un réseau de neurones convolutif appliqué aux frottis sanguins : évaluation critique d’un système d’aide au diagnostic du paludisme déployé au Ngaliema Medical Center

Abstract

Malaria diagnosis in Congolese hospital laboratories still rests on Giemsa-stained blood-film microscopy, a method that is slow, observer-dependent and vulnerable to workload. This article submits a convolutional neural network deployed at the Ngaliema Medical Center in Kinshasa to a full independent re-analysis. The system, built on MobileNetV2 by transfer learning from 5 000 annotated erythrocyte images, was tested prospectively on 150 blood films (75 positive, 75 negative) against a medical biologist reference standard. From the three published indicators we reconstruct the complete confusion matrix, which proves internally exact, then derive the quantities that actually govern clinical use. Sensitivity reaches 96.0% (95% Wilson CI: 88.9-98.6) and specificity 93.3% (85.3-97.1), placing the model at World Health Organization competence level 1 for parasite detection but short of the 95% specificity threshold applied to rapid diagnostic tests, with a negative sample 3.9 times too small to settle the matter. Three findings follow. Diagnostic yield collapses with prevalence: positive predictive value falls from 93.5% at the artificial 50% prevalence of the test set to 61.5% at 10% and 43.1% at 5%, while negative predictive value stays above 99.5% below 10%, making the tool an exclusion instrument rather than a confirmation one. The headline comparison of 4.2 seconds of inference against 20 to 30 minutes of manual reading is inadmissible: automation acts on only 48% of the cycle, and once field acquisition and mandatory human validation are counted the cycle falls from 25.0 to 18.4 minutes, a 26.4% gain and a 35.9% capacity increase, not 99.7%. Finally, recalibrating the threshold to 99.5% sensitivity would release 54.7% of films without expert re-reading at 10% prevalence. At USD 0.30 per examination and USD 6.91 per missed infection averted, value lies in error avoidance and threshold governance, not in raw accuracy.