An AI-based cardiovascular monitoring tool for sepsis identification in ICU

INTRODUCTION Sepsis is one of the major causes of mortality in ICU with an occurrence up to 40% worldwide [1]. It is considered a serious public health issue with 1 in 3 hospitalizations that end in death with sepsis [2]. Sepsis showed a strong influence on cardiovascular functioning in terms of both myocardial and cardiac autonomic dysfunction [3]. The major effects of sepsis on the cardiovascular system can be summarized as follows: a systolic and diastolic cardiac dysfunction, an increased heart rate despite an overall reduction in autonomic modulation of heart activity and an impairment in the baroreflex sensitivity. OBJECTIVES Our study is aimed at exploring the ability of ECG and arterial blood pressure (ABP) waveforms, recorded in the first hour of ICU stay, in recognizing patients with sepsis with an AI-based physiological and cardiovascular monitoring tool. METHODS We extracted the first hour of ECG and ABP waveforms of patients admitted in the ICU, from the publicly available MIMIC-III database on PhysioNet [4,5]. The final population includes 142 patients, 50% of whom with sepsis. The ECG and ABP signals were processed in order to extract the R-peak occurrences from the ECG and the systolic, diastolic and onset occurrences and values from the pressure signal. 68 features were extracted from the heart rate and blood pressure variability domain through mathematical modelling of the closed loop cardiovascular system, which allowed also for the extraction of baroreflex gain [6]. Finally, 7 confoundings comprising the undergoing sedative and vasoactive agent treatment and mechanical ventilation as well as age, gender, diabetes and hypertension were included. A logistic regression model was then trained on a 80% training set and tested on the remaining 20% of data. RESULTS Best results on the test set show an AUROC=0.91 and AUPRC=0.90, thus highlighting the ability of continuously recorded vital signs in recognizing septic patients. Figure 1 depicts the results of the explainability analysis on model decision rules. The model’s strategy results to be consistent with previously summarized clinical considerations about sepsis. Indeed, a lower (impaired) baroreflex gain despite the high systolic pressure values is associated with a higher probability of being septic. Similarly, the second plot shows that an increased heart rate (low RR-interval) is associated with higher sepsis probability and this probability increases when a reduced sympatho-vagal balance (LF/HF) is observed, corresponding to a reduced sympathetic activity or to an overall reduction of autonomic activity. CONCLUSIONS Our results show that continuously recorded patients’ vital signs may help in automatic early identification of sepsis in the early hours of ICU stay. The proposed AI-based physiological monitoring also allows for the extraction of indices able to describe the state of the cardiovascular control system and providing key insights on cardiovascular functioning.