Machine learning based analysis and detection of trend outliers for electromyographic neuromuscular monitoring.

Abstract

Neuromuscular monitoring is frequently plagued by artefacts, which along with the frequent unawareness of the principles of this subtype of monitoring by many clinicians, tends to lead to a cynical attitute by clinicians towards these monitors. As such, the present study aims to derive a feature set and evaluate its discriminative performance for the purpose of Train-of-Four Ratio (TOF-R) outlier analysis during continuous intraoperative EMG-based neuromuscular monitoring. Patient data was sourced from two devices: (1) Datex-Ohmeda Electromyography (EMG) E-NMT: a dataset derived from a prospective observational trial including 136 patients (21,891 TOF-R observations), further subdivided in two based on the type of features included; and (2) TetraGraph: a clinical case repository dataset of 388 patients (97,838 TOF-R observations). The two datasets were combined to create a synthetic set, which included shared features across the two. This process led to the training of four distinct models. The models showed an adequate bias/variance balance, suggesting no overfitting or underfitting. Models 1 and 2 consistently outperformed the others, with the former achieving an F1 score of 0.41 (0.31, 0.50) and an average precision score (95% CI) of 0.48 (0.35, 0.60). A random forest model analysis indicated that engineered TOF-R features were proportionally more influential in model performance than basic features. Engineered TOF-R trend features and the resulting Cost-Sensitive Logistic Regression (CSLR) models provide useful insights and serve as a potential first step towards the automated removal of outliers for neuromuscular monitoring devices. NCT04518761 (clinicaltrials.gov), registered on 19 August 2020.