BACKGROUND: Opioid-induced respiratory depression (OIRD) is traditionally recognized by assessment of respiratory rate, arterial oxygen saturation, end-tidal CO2, and mental status. Although an irregular or ataxic breathing pattern is widely recognized as a manifestation of opioid effects, there is no standardized method for assessing ataxic breathing severity. The purpose of this study was to explore using a machine-learning algorithm for quantifying the severity of opioid-induced ataxic breathing. We hypothesized that domain experts would have high interrater agreement with each other and that a machine-learning algorithm would have high interrater agreement with the domain experts for ataxic breathing severity assessment. METHODS: We administered target-controlled infusions of propofol and remifentanil to 26 healthy volunteers to simulate light sleep and OIRD. Respiration data were collected from respiratory inductance plethysmography (RIP) bands and an intranasal pressure transducer. Three domain experts quantified the severity of ataxic breathing in accordance with a visual scoring template. The Krippendorff alpha, which reports the extent of interrater agreement among N raters, was used to assess agreement among the 3 domain experts. A multiclass support vector machine (SVM) was trained on a subset of the domain expert-labeled data and then used to quantify ataxic breathing severity on the remaining data. The Vanbelle kappa was used to assess the interrater agreement of the machine-learning algorithm with the grouped domain experts. The Vanbelle kappa expands on the Krippendorff alpha by isolating a single rater-in this case, the machine-learning algorithm-and comparing it to a group of raters. Acceptance criteria for both statistical measures were set at >0.8. The SVM was trained and tested using 2 sensor inputs for the breath marks: RIP and intranasal pressure. RESULTS: Krippendorff alpha was 0.93 (95% confidence interval [CI], 0.91-0.95) for the 3 domain experts. Vanbelle kappa was 0.98 (95% CI, 0.96-0.99) for the RIP SVM and 0.96 (0.92-0.98) for the intranasal pressure SVM compared to the domain experts. CONCLUSIONS: We concluded it may be feasible for a machine-learning algorithm to quantify ataxic breathing severity in a manner consistent with a panel of domain experts. This methodology may be helpful in conjunction with traditional measures to identify patients experiencing OIRD.
BACKGROUND: Opioid-induced respiratory depression (OIRD) is traditionally recognized by assessment of respiratory rate, arterial oxygen saturation, end-tidal CO2, and mental status. Although an irregular or ataxic breathing pattern is widely recognized as a manifestation of opioid effects, there is no standardized method for assessing ataxic breathing severity. The purpose of this study was to explore using a machine-learning algorithm for quantifying the severity of opioid-induced ataxic breathing. We hypothesized that domain experts would have high interrater agreement with each other and that a machine-learning algorithm would have high interrater agreement with the domain experts for ataxic breathing severity assessment. METHODS: We administered target-controlled infusions of propofol and remifentanil to 26 healthy volunteers to simulate light sleep and OIRD. Respiration data were collected from respiratory inductance plethysmography (RIP) bands and an intranasal pressure transducer. Three domain experts quantified the severity of ataxic breathing in accordance with a visual scoring template. The Krippendorff alpha, which reports the extent of interrater agreement among N raters, was used to assess agreement among the 3 domain experts. A multiclass support vector machine (SVM) was trained on a subset of the domain expert-labeled data and then used to quantify ataxic breathing severity on the remaining data. The Vanbelle kappa was used to assess the interrater agreement of the machine-learning algorithm with the grouped domain experts. The Vanbelle kappa expands on the Krippendorff alpha by isolating a single rater-in this case, the machine-learning algorithm-and comparing it to a group of raters. Acceptance criteria for both statistical measures were set at >0.8. The SVM was trained and tested using 2 sensor inputs for the breath marks: RIP and intranasal pressure. RESULTS: Krippendorff alpha was 0.93 (95% confidence interval [CI], 0.91-0.95) for the 3 domain experts. Vanbelle kappa was 0.98 (95% CI, 0.96-0.99) for the RIP SVM and 0.96 (0.92-0.98) for the intranasal pressure SVM compared to the domain experts. CONCLUSIONS: We concluded it may be feasible for a machine-learning algorithm to quantify ataxic breathing severity in a manner consistent with a panel of domain experts. This methodology may be helpful in conjunction with traditional measures to identify patients experiencing OIRD.
Authors: Zachary T Glovak; Chelsea Angel; Christopher B O'Brien; Helen A Baghdoyan; Ralph Lydic Journal: Respir Physiol Neurobiol Date: 2021-12-23 Impact factor: 1.931