ICU prognostication: Time to re-evaluate? Register-based studies on improving prognostication for patients admitted to the intensive care unit (ICU)

Abstract

Background: ICU prognostication is difficult because of patients’ prior comorbidities and their varied reasons for admission. The model used for ICU prognostication in Sweden is the Simplified Acute Physiology Score 3 (SAPS 3), which uses information gathered within one hour of ICU admission to predict 30-day mortality. Since the SAPS 3 model was introduced, no biomarkers have been added to it to improve its prognostic performance. For comatose patients admitted to the ICU after cardiac arrest, the prognostication performed after 72 h will either result in the continued observation of the patient or the withdrawal of life-sustaining treatment.
Purpose: 1) To investigate whether adding the biomarker lactate (study I) or high-sensitivity troponin T (hsTnT) (study II) to SAPS 3 adds prognostic value. 2) To investigate whether using a supervised machine learning algorithm called artificial neural networks (ANNs) can improve the prognostic performance of SAPS 3 (study III). 3) To explore whether ANNs can create reliable predictions for comatose patients at the time of hospital admission (study IV) and during the first three days after ICU admission, with or without promising biomarkers (study V).
Methods: 1) To investigate whether the laboratory values of lactate or hsTnT could improve the performance of SAPS 3, we combined patients’ laboratory values on ICU admission at Skåne University Hospital with their SAPS 3 score. 2) Based on all first-time ICU admissions in Sweden from 2009–2017 as retrieved from the Swedish Intensive Care Registry (SIR), we investigated whether ANNs could improve SAPS 3 using the same variables. 3) All out-of-hospital cardiac arrest (OHCA) patients from the Target Temperature Management trial were included for data analysis. Background and prehospital data, along with clinical variables at admission, were used in study IV. Clinical variables from the first three days were used in study V along with different levels of biomarkers defined as clinically accessible (e.g. neuron-specific enolase, or NSE) and research-grade biomarkers (e.g. neurofilament light, or NFL). Patient outcome was the dichotomised Cerebral Performance Category scale (CPC); a CPC of 1–2 was considered a good outcome, and a CPC of 3–5 was considered a poor outcome.
Results: 1) Both lactate and hsTnT were independent SAPS 3 predictors for 30-day mortality in the logistic regression model. In a subgroup analysis, the use of lactate improved the area under the receiver operating characteristic curve (AUROC) for cardiac arrest and septic patients, and the use of hsTnT improved the AUROC for septic patients. 2) The overall performance of the SAPS 3 model in Sweden was improved by the use of ANNs. Both the discrimination (AUROC 0.89 vs 0.85, p < 0.001) and the calibration were improved when the two models were compared on a separate test set (n = 36,214). 3) An ANN model outperformed a logistic-regression-based model in predicting poor outcome on hospital admission for OHCA patients. Incorporating biomarkers such as NSE improved the AUROC over the course of the first three days of the ICU stay; when NFL was incorporated, the prognostic performance was excellent from day 1.
Conclusion: Lactate and hsTnT probably add prognostic value to SAPS 3 for patients admitted to the ICU with sepsis or after cardiac arrest (lactate only). An ANN model was found to be superior to the SAPS 3 model (Swedish modification) and corrected better for age than SAPS 3. A simplified ANN model with eight variables showed performance similar to that of the SAPS 3 model. For comatose OHCA patients, an ANN model improved the accuracy of the prediction of the long-term neurological outcome at hospital admission. Furthermore, when it used cumulative information from the first three days after ICU admission, an ANN model showed promising prognostic performance on day 3 when it incorporated clinically accessible biomarkers such as NSE, and it showed promising performance on days 1–3 when it incorporated research-grade biomarkers such as NFL.

• author: Andersson, Peder
• supervisor: Attila Frigyesi; Jonas Björk
• opponent: Consultant, clinical associate professor, PhD Christiansen, Christian Fynbo, Department of Clinical Epidemiology (DCE), Aarhus University, Denmark
• organization: Anesthesiology and Intensive Care; Mathematical Statistics; eSSENCE: The e-Science Collaboration
• publishing date: 2021
• type: Thesis
• publication status: published
• subject: Anesthesiology and Intensive Care
• keywords: Intensive Care, Critical Care, Mortality, Cerebral Performance Category, Cardiac arrest, Out-of-hospital cardiac arrest, Sepsis, Lactate, Troponin, Neuron-specific enolase, Neurofilament light, Age, Elderly, Prognostication, Prediction, Scoring system, Neural net, Artificial neural network, Deep learning, Artificial intelligence
• in: Lund University, Faculty of Medicine Doctoral Dissertation Series
• issue: 2021:11
• pages: 84 pages
• defense location: Konferensrummet, Intensiv- och perioperativ vård, arbetsavdelningen, Centralblocket, hisshall B, plan 6, SUS, Lund.
• defense date: 2021-02-11 13:30:00
• ISSN: 1652-8220
• ISBN: 978-91-8021-017-1
• language: English
• LU publication?: yes
• id: 7c6b4264-b943-454d-8515-fec999031cbc
• date added to LUP: 2021-01-18 00:50:51
• date last changed: 2022-06-27 14:27:53