researchsquare.comSource publication:
Loading content…
Sepsis is often flagged at emergency triage, but a neural network returns only a probability that clinicians cannot check. ShAprox fits each feature's Shapley value to a curve or mean and adds them to a base rate, so the prediction is a readable sum. Across 28,072 emergency attendances at two NHS trusts, ShAprox reached 0.824 external-test AUC, comparable to the neural network's 0.818, with training and internal-test performance closer together.
A sepsis model can estimate the probability of a later sepsis diagnosis from the first vital signs, demographics and comorbidities recorded when a patient reaches emergency triage, and that probability is built by adding one contribution per feature so a clinician can inspect each term. Deployed sepsis alerts such as NEWS2 and the St. John Sepsis Agent are expert systems that score by preset rules; machine learning can predict earlier, but explanation tools such as SHAP and LIME did not give clinicians enough insight to trust a black-box output. Across 28,072 patient spells at two acute NHS trusts, ShAprox reached 0.824 external-test AUC and 0.906 SAS, against 0.818 and 0.911 for the neural network NNS1; on internal testing the figures were 0.735 and 0.807 versus 0.756 and 0.814.
The prediction equation adds an approximate Shapley value function for each feature to a base rate, using a third-degree curve of best fit for continuous features and a mean for binary ones, so the same blood-pressure reading contributes equally for any patient and a small change shifts the prediction only slightly. Shapley values come from game theory, distributing a payoff among players by contribution; SHAP applied them as per-case post-hoc attributions rather than as the predictive model itself. Training used 80% of Site 1 data, internal testing the remaining 20%, and all 14,978 Site 2 spells formed the external test; the network took 55 input variables, 40% dropout and three hidden layers over 106 epochs.
The results cover adults admitted through the emergency department between March 2019 and August 2020 at two acute NHS trusts whose diagnosis appears on the suspicion-of-sepsis list, totaling 28,072 patient spells. The model uses only first-recorded vital signs, age, gender, ethnicity and Elixhauser comorbidities, not blood test results. The authors state it should be retrained and validated on all emergency patients before use in an emergency ward.
The dataset is limited to suspicion-of-sepsis patients, and the authors believe non-sepsis patients would have noticeably distinct vitals and characteristics, though this is uncertain. Missing vitals were more frequent among non-sepsis patients and were filled with random healthy values. The authors also state that Pakistani ethnicity lowering the prediction should not be read causally and may reflect confounding or coding differences at Site 1.