Author: Vishal Nangalia
Acute kidney injury (AKI) is both prevalent (ranging from 14 - 21%) and results in severe mortality (23.3%) in hospitalised patients. However, over 50% of poor outcome from AKI is deemed preventable. Thus, worldwide initiatives have focussed on the early identification and management of AKI. However, the current AKI severity staging for predicting poor outcome leaves a lot to be desired, poor- sensitivity, specificity and positive predictive value. This study explored whether an advanced Machine Learning (ML) approach could better predict outcome from AKI, earlier in its disease trajectory. Specifically, this study examined the prevalence of AKI in the largest dataset yet collated of hospitalised patients, from over fourteen UK National Health Service Trusts, comprising over twenty hospitals, collected over a period from 2005 to 2015. All adult patients who were admitted to hospital and had at least one serum creatinine measurement in hospital were included. The NHS England AKI (NHSE-AKI) algorithm was used to ascertain the prevalence of AKI and the severity stages of AKI among all the admissions. Both the first AKI stage on admission (AKIfirst) and the maximum (AKImax) reached were calculated. Using additional, yet routinely collected variables of: additional blood tests, administrative data, and key co-morbidities, a machine learning model (ML-AKI) was created applying gradient boosted machines to predict outcome from AKI, specifically whether a patient died or required renal replacement therapy in-hospital (D-RRT). Positive predictive value thresholds ranging from 1:2 (ML50) to 1:3 (ML33) were also calculated for the model’s ability to predict D-RRT. Of the 1,972,130 admissions analysed, 170,596 (8.6%) developed AKI. Community acquired AKI was present in 65,722 (38.6%) of these AKI admissions. AKI was twice as common in emergency vs planned admissions (10.7% (122,346) vs 5.8% (48,240), p<1 x 10-10). D-RRT occurred in 21% (35,832) of those who developed AKI, compared with 1.6% (28,464: died=26,029; RRT=2,577) (p<0.001) in those who did not. Such poor outcome in AKI admissions was more prevalent in emergency than planned admissions (25.1% (30,726) vs 10.6% (5,106) respectively, p<1 x 10-5). D-RRT risk was related to higher AKImax and AKIfirst stages, with D-RRT rates ranging from 13.6% to 42.4% for AKImax and 18.7% to 32.8% for AKIfirst Overall, predictive performance of the ML-AKI model exceeded that of the NHSE-AKI algorithm. It achieved an area under the receiver operator curve of 85.9% (with D-RRT was coded as 0/1: log loss: 0.36, mean squared error: 0.11, and mean per class error: 0.24) on the test dataset. The rate of D-RRT for ML33 (33.3%) and ML50 (50%) was higher than for any AKIfirst stage (18.7 to 32.8%). Even when merging AKI stages 2 and 3 together, their performance fell short of both ML33 and ML50 for D-RRT (AKIfirst 2&3 = 28.8%) and sensitivity (AKIfirst 2&3 = 31.5%, ML33 = 95.3%, ML50 = 74.3%). Thus, I have 1) highlighted weaknesses in the current AKI detection algorithms and 2) demonstrated a more accurate alternative using machine learning to better target escalation of care when compared with the existing stages of the AKI algorithm.
Funding Acknowledgement (If Applicable): Medical Research Council, NIHR UCLH Biomedical Research Institute