Work

Automatic Segmentation and Classification of Cervical Accelerometry Signals: Towards the Instrumental Detection of Unsafe Swallowing

Year: 2017
Type: dissertation
Abstract: Dual-axis cervical accelerometry has been proposed as a method for non-invasive clinical evaluation of swallowing function in people with risk for dysphagia (swallowing impairment). Acceleration signals in anterior-posterior and superior-inferior anatomical directions are processed and automatically segmented for use in classification experiments. However, automatic segmentation is often too liberal, admitting both pre- and post-swallowing activity and non-swallow segments. These segmentation shortcomings adversely affect feature extraction and ultimately classification of swallowing function. We propose a kernel density estimation-based algorithm to adaptively trim the swallow segments, and energy and noise-floor algorithms to reduce the number of false positive swallow segments. Dramatic reductions (-85.4%) in false positives can be achieved with a moderate loss of true positives (-15.1%). Furthermore, we propose a detection system for instructed and reflexive coughs, discriminating accelerometry signals associated with coughs from those representing swallows, tongue movements, and speech. Using binary genetic feature selection and a support vector machine, the proposed system achieved a cough detection accuracy of 99.26±0.12% when discriminating between instructed cough and rest state accelerometry signals. An accuracy of 90±13.9% was achieved using elastic net feature selection and a support vector machine when classifying reflexive coughs and rest signals. When discriminating instructed and reflexive coughs from all other non-cough artefacts, the proposed system achieved accuracies of 90.2±3.6% and 80.3±10.5%, respectively. Additionally, all previous cervical accelerometry studies have relied on videofluoroscopy-demarcated segments rather than the automatic localization of segments of interest, which precludes direct translation of these algorithms into a stand-alone tool. Moreover, previous studies have adopted an ipsative classification paradigm, wherein, for every unseen case, the classifier must make a forced choice (e.g. between safe and unsafe swallowing). To address these shortcomings, bolus length estimation and instance selection were introduced as enhancements to swallowing accelerometry classification, on one hand liberating classification algorithms from manual segmentation of swallows and secondly affording the classifier the freedom to abstain from a decision in the face of uncertainty. We demonstrated that together these enhancements lead to an improvement in area under the curve (83.6±5.7%) in the discrimination between safe and unsafe swallows in a sizable clinical data set.
Language: English
Open Access status: green
OpenAlex ID: W3088441124
Cites: 156
Cited by:
Related to: 9