[Paper Review] Machine Learning based Anomaly Detection for Smart Shirt: A Systematic Review

E. C. Nunes|arXiv (Cornell University)|Mar 7, 2022

Anomaly Detection Techniques and Applications4 citations

TL;DR

This systematic review analyzes 11 studies (2017–2021) on machine learning-based anomaly detection in smart shirts, focusing on fall detection, cardiac arrhythmia, and posture abnormalities. It finds that supervised learning, especially Support Vector Machines (SVM), achieves over 90% accuracy using sensor data from devices like Hexoskin, with fall detection being the most common anomaly type.

ABSTRACT

In recent years, the popularity and use of Artificial Intelligence (AI) and large investments on theInternet of Medical Things (IoMT) will be common to use products such as smart socks, smartpants, and smart shirts. These products are known as Smart Textile or E-textile, which has theability to monitor and collect signals that our body emits. These signals make it possible to extractanomalous components using Machine Learning (ML) techniques that play an essential role in thisarea. This study presents a Systematic Review of the Literature (SLR) on Anomaly Detection usingML techniques in Smart Shirt. The objectives of the SLR are: (i) to identify what type of anomalythe smart shirt; (ii) what ML techniques are being used; (iii) which datasets are being used; (iv)identify smart shirt or signal acquisition devices; (v) list the performance metrics used to evaluatethe ML model; (vi) the results of the techniques in general; (vii) types of ML algorithms are beingapplied.The SLR selected 11 primary studies published between 2017-2021. The results showed that6 types of anomalies were identified, with the Fall anomaly being the most cited. The Support VectorMachines (SVM) algorithm is most used. Most of the primary studies used public or private datasets.The Hexoskin smart shirt was most cited. The most used metric performance was Accuracy. Onaverage, almost all primary studies presented a result above 90%, and all primary studies used theSupervisioned type of ML.

Motivation & Objective

• To identify the types of anomalies detected in smart shirts using machine learning.
• To analyze the machine learning techniques, datasets, and performance metrics used in existing research.
• To evaluate the performance of ML models in detecting physiological and behavioral anomalies in smart shirts.
• To determine the most frequently used smart shirt or signal acquisition devices in current studies.
• To assess the state of the art in supervised machine learning for real-time anomaly detection in wearable textiles.

Proposed method

• Conducted a systematic literature review (SLR) following Kitchenham and Charters' methodology to identify relevant primary studies (2017–2021).
• Selected 11 high-quality primary studies based on predefined inclusion and exclusion criteria.
• Extracted data on anomaly types, ML algorithms, datasets, signal acquisition devices, performance metrics, and model results.
• Classified ML techniques as supervised, with a focus on classification performance for anomaly detection.
• Evaluated model performance using standard metrics such as accuracy, precision, recall, F1-score, and specificity.
• Mapped results to identify trends in algorithm popularity, sensor types (e.g., accelerometer, ECG), and device models (e.g., Hexoskin).

Experimental results

Research questions

• RQ1 What types of anomalies are detected in smart shirts using machine learning?
• RQ2 Which machine learning techniques are most commonly applied for anomaly detection in smart shirts?
• RQ3 What datasets are used to train and evaluate the ML models in this domain?
• RQ4 Which smart shirt or signal acquisition devices are most frequently used in the studies?
• RQ5 What performance metrics are used, and how do the models perform in terms of accuracy and classification metrics?

Key findings

• Six types of anomalies were identified, with 'fall' being the most frequently studied, appearing in five of the 11 primary studies.
• Support Vector Machines (SVM) were the most widely used ML algorithm, applied in seven of the 11 studies.
• All studies used supervised learning, with model accuracy exceeding 90% in nearly all cases—ranging from 90.01% to 98.5%.
• The Hexoskin smart shirt was the most cited device for data acquisition, used in multiple studies.
• Accuracy was the most commonly reported performance metric, used in eight studies, followed by F1-score, precision, and recall.
• The most effective models achieved over 98% accuracy (e.g., SVM in A1 and A11), with high specificity (98.5%) and sensitivity (97.6%) in some cases.