Sumitra Mukhopadhyay

Continuous biometric authentication employs several security measures, which endorse the legitimate user in a continuous manner with consideration of issues such as security, safety, and convenience. However, the reported state-of-the-art works put less focus on the validation of authentication model with unseen unauthorized users and hardware implementation. In this letter, a real-time 2-stage photoplethysmogram (PPG) assisted continuous authentication technique is presented with legitimate user authorization (anomaly detection) followed by identification. A denoised short segment of 5 s PPG signal is used to extract time, frequency, entropy, and statistical features followed by recursive feature elimination, XGBOD, and RF-based authorization and identification, respectively. The proposed technique was tested with PhysioNet BIDMC PPG and Respiration dataset (53 subjects) and laboratory collected dataset (10 subjects) to detect authorized users and unauthorized-unseen users with overall accuracy and F1 score of 96.62% and 96.63%, respectively, for real-time standalone implementation. The end-to-end model was deployed in a BCM2837B0, Cortex-A53 64-bit SoC (1.4 GHz processor, 1 GB RAM) with a latency of ∼8.038 s and memory consumption of ∼49.891 MB for 5 s PPG data. The present technique may be extended to further resource constrained environment with increased number of authorizations.

Automatic cardiac arrhythmia detection (CAD) from ECG is a challenging topic for biomedical research community. In recent times, machine learning and deep learning models have been widely adopted to achieve good prediction accuracy. However, the primary obstacle towards their standalone implementation are often high complexity and memory requirement, preventing low resource embedding. The other one is low prediction accuracy towards minority classes. This work presents a new data augmentation model, a combination of borderline majority class undersampling (BMU) and chaos-based minority class oversampling (CMO). In BMU the undersampling was done with priority to the majority class instances which are far away from the boundary. In CMO, the beat synthesis was done by adaptive chaos addition to the parent beats of minority classes so that they remain clinically acceptable. A simple classification flow using an autoencoder with random forest was used with the synthesized data for multiclass CAD. For performance evaluation three benchmark datasets, viz., MITDB, INCART and SVDB datasets were used. It was found that the F1 score of the minority classes have increased, with reduction in false positive and negative rates as compared to few published works employing complex architectures. Also, the simple CAD model could achieve equally good accuracy with much lower (70%) latency, and lower (90%) memory requirement than few reported works using MITDB dataset in a standalone hardware implementation (ARM v6 controller). This research shows great promise towards developing low-complexity wearable CAD monitors.

Computerized myocardial infarction (MI) detection and localization can be useful for early prevention of its aggravation and related cardiac health complications. However, the published research either focuses on binary classification or implements complex classifiers for localization to achieve good accuracy. In this work, the objective is to implement an 11-class MI localization system on resource-constrained hardware with low complexity and latency. A simple and optimized autoencoder- k -NN classifier has been used to achieve accuracy and F1-score of 99.74% and 99.20% respectively while evaluating single lead ECG features from the PTBDB dataset. A standalone hardware implementation with an ARM-v6-based controller resulted in a latency and runtime memory engagement of 0.48s and 4.31 MB respectively to process 5 s ECG data. The present research can be useful for quick screening of MI for portable healthcare applications.

Early automatic detection of cardiac arrhythmia in an ambulatory health monitoring setup can reduce the mortality rates. The existing approaches on multiclass arrhythmia detection (MCAD) techniques, adopting either time domain morphological features (TDMF) fail to extract hidden pattern from the single lead ECG resulting in poor accuracy (Ac); or, adopting deep-learning (DL) techniques become too complex or heavy in architecture, hindering their easy implementation in low resource architecture for wearable applications. This work presents a judicious hybridization of both approaches in an end-to-end MCAD using 2-stage classifier to achieve higher accuracy and low latency while detecting 15 classes of cardiac rhythms (14 types of arrhythmia beats or normal sinus rhythms). The memory resource saving is achieved through the beat skeletonization process and two-stage feature minimization. The stage#1 classifier uses 14 TDMF features fed to a random forest (RF) regressor and provides probability values for normal sinus rhythm, atrial or ventricular conduction problem. The stage#2 classifier, another RF, accepts stacked autoencoder features, TDMF features and output of stage#1 to achieve a blind test Ac of 99.21% using only 13 features while evaluated with full mitdb records. A standalone on-board application of the 2-stage MCAD using ARM v6 single core microcontroller yields a detection latency of 782 ms and consumes 390 kbytes of memory per cardiac cycle, proving a feasible solution for wearable cardiac monitors. It also provides competitive performance with state-of-the-art methods, and outperforms the limited available research on hardware implementation in terms of number of classes handled and Ac.