OBJECTIVE: We aimed to evaluate the validity of a MARSIPAN-guidance-adapted Early Warning System (MARSI MEWS) and compare it to the National Early Warning Score (NEWS) and an adapted version of the Physical Risk in Eating Disorders Index (PREDIX), to ascertain whether current practice is comparable to best-practice standards. METHODS: We collated 3,937 observations from 36 inpatients from Addenbrookes Hospital over 2017-2018 and used three independent raters to create a "gold standard" of deteriorating cases. We ascertained performance metrics (Receiver Operating Characteristic Area Under the curve) for MARSI MEWS, NEWS and PREDIX; we also tested the proof of concept of a machine-learning-based early-warning-system (ML-EWS) using cross-validation and out-of-sample prediction of cases. RESULTS: The MARSI MEWS system showed higher ROC AUC (0.916) compared to NEWS (0.828) or PREDIX (0.865). ML-EWS (random forest) performed well at independent samples analysis (0.980) and multilevel analysis (0.922). CONCLUSION: MARSI MEWS seems most suitable for identifying critically deteriorating cases in anorexia nervosa inpatient population. We did not examine community practice in which the PREDIX arguably remains the best to ascertain deteriorating cases. Our results also provide a first proof of concept for the development of artificial-intelligence-based early warning systems in anorexia nervosa. Implications for inpatient clinical practice in eating disorders are discussed.
OBJECTIVE: We aimed to evaluate the validity of a MARSIPAN-guidance-adapted Early Warning System (MARSI MEWS) and compare it to the National Early Warning Score (NEWS) and an adapted version of the Physical Risk in Eating Disorders Index (PREDIX), to ascertain whether current practice is comparable to best-practice standards. METHODS: We collated 3,937 observations from 36 inpatients from Addenbrookes Hospital over 2017-2018 and used three independent raters to create a "gold standard" of deteriorating cases. We ascertained performance metrics (Receiver Operating Characteristic Area Under the curve) for MARSI MEWS, NEWS and PREDIX; we also tested the proof of concept of a machine-learning-based early-warning-system (ML-EWS) using cross-validation and out-of-sample prediction of cases. RESULTS: The MARSI MEWS system showed higher ROC AUC (0.916) compared to NEWS (0.828) or PREDIX (0.865). ML-EWS (random forest) performed well at independent samples analysis (0.980) and multilevel analysis (0.922). CONCLUSION: MARSI MEWS seems most suitable for identifying critically deteriorating cases in anorexia nervosa inpatient population. We did not examine community practice in which the PREDIX arguably remains the best to ascertain deteriorating cases. Our results also provide a first proof of concept for the development of artificial-intelligence-based early warning systems in anorexia nervosa. Implications for inpatient clinical practice in eating disorders are discussed.