Photo Credit: iStock.com/ClaudioVentrella

Deep learning using routine lab data showed promise in distinguishing candidemia from bacteremia but fell short of outperforming established biomarkers.  

Researchers conducted a retrospective study published in June 2025 issue of Infectious Diseases and Therapy to assess the impact of early diagnosis on outcomes in individuals with candidemia, bloodstream infection (BSI) particularly those presenting with septic shock.  

They evaluated the effectiveness of a deep learning model in distinguishing candidemia from bacteremia at an early stage. The model was trained using a large dataset composed of automatically extracted laboratory variables. The approach aimed to support early differential diagnosis based on routine lab data.  

The results showed that out of 12,483 episodes, 1,275 (10%) were candidemia, and 11,208 (90%) were bacteremia. On the training set, the deep learning model achieved a sensitivity of 0.80, specificity of 0.59, positive predictive value (PPV) of 0.18, weighted PPV (wPPV) of 0.88, and negative predictive value (NPV) of 0.96 with an area under the curve (AUC) of 0.69. On the test set, sensitivity was 0.70, specificity 0.58, PPV 0.16, wPPV 0.87, and NPV 0.95 with an AUC of 0.64. The model’s performance was further evaluated in a subgroup with serum β-d-glucan (BDG) and procalcitonin (PCT), where feature selection and transfer learning failed to improve diagnostic accuracy beyond that of BDG and PCT alone.  

Investigators concluded that a deep learning model trained on nonspecific laboratory features showed limited added diagnostic value over specific markers for distinguishing candidemia from bacteremia, though it demonstrated potential for future integration with clinical data. 

Source: link.springer.com/article/10.1007/s40121-025-01171-w