Photo credit: istock.com/claudioventrella
Deep learning using regular lab data showed promising in distinguishing candidemia from bacteremia, but was not lacking in the performance of established biomarkers.
The researchers conducted a retrospective study of infectious diseases and treatment in June 2025 to assess the impact of early diagnosis on candidemia outcomes. Blood flow Infection (BSI) especially those who present with septic shock.
They evaluated the effectiveness of the deep learning model in distinguishing candidemia from bacteremia in the early stages. This model was trained using a large dataset consisting of automatically extracted laboratory variables. This approach is intended to support early differential diagnosis based on routine lab data.
Results showed that out of 12,483 episodes, 1,275 (10%) had candidemia and 11,208 (90%) had bacteremia. In the training set, the deep learning model achieved a sensitivity of 0.80, a specificity of 0.59, a positive predicted value of 0.18 (PPV), a weighted PPV of 0.88 (WPPV), and a negative predicted value of 0.96 (NPV) of 0.69. In the test set, sensitivity was 0.70, specificity was 0.58, PPV 0.16, WPPV 0.87, and NPV 0.95, and AUC was 0.64. Model performance was further evaluated in subgroups containing serum β-D-glucan (BDG) and procatonin (PCT). Here, feature selection and transfer learning failed to improve diagnostic accuracy beyond BDG and PCT only.
The researchers concluded that deep learning models trained in non-specific laboratory features showed potential for future integration with clinical data, but showed limited diagnostic values than specific markers for distinguishing candidemia from bacteremia.
sauce: link.springer.com/article/10.1007/S40121-025-01171-W
