Newborn Cry Acoustics in the Assessment of Neonatal Opioid Withdrawal Syndrome Using Machine Learning

Abstract

Importance: The assessment of opioid withdrawal in the neonate, or neonatal opioid withdrawal syndrome (NOWS), is problematic because current assessment methods are based on subjective observer ratings. Crying is a distinctive component of NOWS assessment tools and can be measured objectively using acoustic analysis.

Objective: To evaluate the feasibility of using newborn cry acoustics (acoustics referring to the physical properties of sound) as an objective biobehavioral marker of NOWS.

Design, setting, and participants: This prospective controlled cohort study assessed whether acoustic analysis of neonate cries could predict which infants would receive pharmacological treatment for NOWS. A total of 177 full-term neonates exposed and not exposed to opioids were recruited from Women & Infants Hospital of Rhode Island between August 8, 2016, and March 18, 2020. Cry recordings were processed for 118 neonates, and 65 neonates were included in the final analyses. Neonates exposed to opioids were monitored for signs of NOWS using the Finnegan Neonatal Abstinence Scoring Tool administered every 3 hours as part of a 5-day observation period during which audio was recorded continuously to capture crying. Crying of healthy neonates was recorded before hospital discharge during routine handling (eg, diaper changes).

Exposures: The primary exposure was prenatal opioid exposure as determined by maternal receipt of medication-assisted treatment with methadone or buprenorphine.

Main outcomes and measures: Neonates were stratified by prenatal opioid exposure and receipt of pharmacological treatment for NOWS before discharge from the hospital. In total, 775 hours of audio were collected and trimmed into 2.5 hours of usable cries, then acoustically analyzed (using 2 separate acoustic analyzers). Cross-validated supervised machine learning methods (combining the Boruta algorithm and a random forest classifier) were used to identify relevant acoustic parameters and predict pharmacological treatment for NOWS.

Results: Final analyses included 65 neonates (mean [SD] gestational age at birth, 36.6 [1.1] weeks; 36 [55.4%] female; 50 [76.9%] White) with usable cry recordings. Of those, 19 neonates received pharmacological treatment for NOWS, 7 neonates were exposed to opioids but did not receive pharmacological treatment for NOWS, and 39 healthy neonates were not exposed to opioids. The mean of the predictions of random forest classifiers predicted receipt of pharmacological treatment for NOWS with high diagnostic accuracy (area under the curve, 0.90 [95% CI, 0.83-0.98]; accuracy, 0.85 [95% CI, 0.74-0.92]; sensitivity, 0.89 [95% CI, 0.67-0.99]; specificity, 0.83 [95% CI, 0.69-0.92]).

Conclusions and relevance: In this study, newborn acoustic cry analysis had potential as an objective measure of opioid withdrawal. These findings suggest that acoustic cry analysis using machine learning could improve the assessment, diagnosis, and management of NOWS and facilitate standardized care for these infants.

Figure 1.. Model Stacking Diagram

NOWS indicates neonatal opioid withdrawal syndrome.

Figure 2.. Flowchart of Data Reduction Strategy

NOWS indicates neonatal opioid withdrawal syndrome.

Figure 3.. Acoustic Feature Importance and Stability Plots

Mean importance estimates represent the relative contribution of a given variable to the final prediction across iterations of the cross-validation procedure. Variable stability was estimated by examining how frequently a given variable was retained for analysis during the feature selection step; for stability estimates, 100% indicates that a given variable was retained during all iterations of the cross-validation procedure. Both high importance and high stability are desirable.

Figure 4.. Receiver Operating Characteristic Curve and Precision-Recall Curve of the Final Prediction

The triangle symbol indicates the point that was closest to perfect prediction (ie, values of 1 for sensitivity and 1 for specificity). In the precision-recall curve, precision is equivalent to positive predictive value, and recall is equivalent to sensitivity. AUC indicates area under the curve; ROC, receiver operating characteristic.