Prediction of Late-Onset Sepsis in Preterm Infants Using Monitoring Signals and Machine Learning

Abstract

Objectives: Prediction of late-onset sepsis (onset beyond day 3 of life) in preterm infants, based on multiple patient monitoring signals 24 hours before onset.

Design: Continuous high-resolution electrocardiogram and respiration (chest impedance) data from the monitoring signals were extracted and used to create time-interval features representing heart rate variability, respiration, and body motion. For each infant with a blood culture-proven late-onset sepsis, a Cultures, Resuscitation, and Antibiotics Started Here moment was defined. The Cultures, Resuscitation, and Antibiotics Started Here moment served as an anchor point for the prediction analysis. In the group with controls (C), an "equivalent crash moment" was calculated as anchor point, based on comparable gestational and postnatal age. Three common machine learning approaches (logistic regressor, naive Bayes, and nearest mean classifier) were used to binary classify samples of late-onset sepsis from C. For training and evaluation of the three classifiers, a leave-k-subjects-out cross-validation was used.

Setting: Level III neonatal ICU.

Patients: The patient population consisted of 32 premature infants with sepsis and 32 age-matched control patients.

Interventions: No interventions were performed.

Measurements and main results: For the interval features representing heart rate variability, respiration, and body motion, differences between late-onset sepsis and C were visible up to 5 hours preceding the Cultures, Resuscitation, and Antibiotics Started Here moment. Using a combination of all features, classification of late-onset sepsis and C showed a mean accuracy of 0.79 ± 0.12 and mean precision rate of 0.82 ± 0.18 3 hours before the onset of sepsis.

Conclusions: Information from routine patient monitoring can be used to predict sepsis. Specifically, this study shows that a combination of electrocardiogram-based, respiration-based, and motion-based features enables the prediction of late-onset sepsis hours before the clinical crash moment.

Keywords: infant; intensive care units; machine learning; monitoring; neonatal; physiologic; predictive value of tests; premature; sepsis.

Figure 1.

Illustration of the Cultures, Resuscitation, and Antibiotics Started Here (CRASH) moment and equivalent crash moment (ECM) after matching. LOS = late-onset sepsis.

Figure 2.

Time series analysis for late-onset sepsis (LOS) (red) and controls (green) for 32 matched patients (Online Figure I, Supplemental Digital Content, http://links.lww.com/CCX/A460; 62 and 69 patients of the total cohorts before matching). The value displayed at each timepoint is the average value of the last hour preceding that timepoint. * and ** correspond to a significant difference between LOS and with p < 0.05 and p < 0.01, respectively. As the greatest differences for various features were observed approximately 3 hr before Cultures, Resuscitation, and Antibiotics Started Here (CRASH), 3-hr segments were used for training the classifiers. ECG = electrocardiogram, pDec = percentage of decelerations, RespIDR = interdecile range of respiratory rate, RespSD = the sd of respiratory rate, RMSSD = square root of the mean of the squares of successive differences between adjacent normal-to-normal intervals, SDDec = the sd of RR interval corresponding to percentage of decelerations, SDNN = sd of the RR interval, SII = signal instability index, SII-IDR = interdecile range of the SII, SII-Skew = skewness of the SII.

Figure 3.

Accuracy per time interval of 1 hr, based on the machine learning model trained using 3-hr segment. Results for: A, logistic regressor. B, Naive Bayes classifier. C, Nearest mean classifier. In general, the accuracy of combining all features is superior than the accuracy of a single feature. CRASH = Cultures, Resuscitation, and Antibiotics Started Here, HRV = heart rate variability.