Fig. 2: Case studies: Evaluating KnowRare’s adaptability and generalisation in clinical practice. | npj Digital Medicine

Fig. 2: Case studies: Evaluating KnowRare’s adaptability and generalisation in clinical practice.

From: Bridging data gaps of rare conditions in ICU: a multi-disease adaptation approach for clinical prediction

Fig. 2

Studies include: a Analysis of required source conditions: Assessing KnowRare’s adaptability to different hospital datasets and clinical prediction tasks by varying the proportion of source conditions included from 1% to 100%. b Impact of condition KG sparsity: Assessing KnowRare’s sensitivity and adaptability to varying KG completeness by retaining different proportions of top-weighted edges (from 1% to 100%) in the KG. c Generalisation to common conditions under limited-data scenarios: Evaluating KnowRare’s robustness by training it with only 10% of the available septicemia data, while the LSTM baseline model uses septicemia data ranging from 10% to 100%. Rows correspond to the five prediction tasks: (1) 90-day mortality prediction after hospital discharge (MIMIC-III), (2) 30-day readmission prediction after hospital discharge (MIMIC-III), (3) ICU mortality prediction (eICU), (4) Remaining length of stay prediction (eICU), and (5) Phenotyping prediction (eICU). Points represent mean values, and shaded regions indicate standard deviation over five runs. Plots are generated using matplotlib (Python).

Back to article page