<p dir="ltr">Malaria is a significant global health challenge, with effective control and eradication of malaria depend heavily on testing all suspected cases; however, school-age children are frequently overlooked, with potential for severe complications. Recent advances in retinal imaging and computer vision algorithms have shown potential for improving malaria detection. Despite these advancements, non-invasive point-of-care malaria detection has yet to be realized, primarily due to the need for specialized equipment. We investigate state-of-the-art deep learning models for automated prediction of malaria risk in a non-invasive manner. Specifically, we utilize convolutional neural networks - ResNet-18 and VGG-11 - for classifying malaria presence using photographs of the bulbar and palpebral conjunctiva, which are easily accessible sites and can contain visible symptoms for malaria. The dataset comprises 4,220 photographs from 405 children aged 5–15 years, collected using various smartphone models in a high-malaria region of Rwanda. We incorporate transfer learning with pretrained weights, green channel extraction, histogram equalization, and data augmentation, with the aim of enhancing model performance. Our results show that transfer learning and data augmentation at certain hyperparameters improve model stability and performance with ResNet-18 and VGG-11 based on quantitative metrics like recall, F1-Scores, and test accuracies. Future research focusing on color correction and more advanced model training approaches may overcome limitations and further enhance diagnostic ability of using photographs. This approach offers a rapid, non-invasive, and computationally efficient solution for malaria risk prediction, potentially advancing malaria detection in resource-limited settings.</p>