Description:
Only positive instances of various events, e.g., secondary diagnoses, are actively labeled in hospital administrative data. In line with this, several studies indicate underreporting of adverse events such as sepsis. The gold standard for relabeling of uncoded sepsis cases, medical record review, is laborious, costly, and infeasible to execute for identifying sepsis in large, national datasets. We apply a positive unlabeled (PU) learner as a novel approach to identify sepsis cases from hospital administrative data. We exploit the Hospital Case Cost Statistic from the Swiss Federal Statistics Office (data years 2017 to 2019) including 72 cost attributes at case level. We hypothesize that these cost data should prove effective for learning a classification model as positive sepsis cases in the unlabeled data should exhibit similar cost patterns as labeled positive examples. We randomly draw 200,000 unlabeled examples from the full dataset and add 64,915 positive examples of sepsis labeled in the observation period for model training and evaluation. We train a robust PU learner proven in other applications, AdaSampling, with support vector machine as classification model. For model evaluation, we perform five-fold crossvalidation. Due to the PU setting, we can only use positive examples in the test set and estimate recall along with precision and recall at 10%, 20%, and 30% for four different evaluation scenarios, changing the coding strategy for labeling sepsis cases. Our model has a recall of 85.1% when labeling sepsis cases explicitly in the test set. Recall decreases to 55.5% when labeling sepsis cases exclusively with an implicit coding strategy. Recall at k% is highest for the evaluation scenarios focusing on implicit coding strategies, yet remains relatively low throughout all scenarios. Precision at k% is highest when only considering cases as positive examples that would be labeled according to both the explicit as well as implicit coding strategy (e.g., 92.3% for k=10%). Compared to the sensitivity of directly identifying sepsis cases from hospital administrative data reported in studies using medical record review, the recall of our model is high. We propose a two step process using PU learning for increasing the quality of hospital administrative data and performing sensitity analyses for health economic and health services research.