Ashish Kaushal

Heart Rate Estimation

• Developed a deep learning approach that combines heart rate (HR) estimation with confidence assessment, leveraging STFT-based BCG spectrograms
• Utilized a Depthwise Separable Convolutional Neural Network (DW-CNN) to capture heart rate patterns in both frequency and time domains
• Introduced a confidence estimation framework built on a pre-trained CNN-based HR regression model, comprising an Encoder (for STFT and Piezo features), a Predictor (for HR output), and a Confidence Model to assess reliability
• Achieved a Mean Absolute Error (MAE) of 2.50 bpm and a detection rate (DR) of 83.90% across 70,211 reference points with 58,908 valid HR measurements

Sleep Apnea Detection

• Implemented a deep learning model for sleep apnea detection using Piezoelectric (BCG), SpO2, and FFT features, integrating multi-branch CNNs, attention mechanisms, and an LSTM for sequence modeling
• Designed three parallel CNN pipelines for Piezo, SpO2, and FFT, capturing distinct feature sets prior to attention-based fusion
• Employed multi-head attention to enhance cross-modal interactions and integrated a bidirectional LSTM for sequential dependency learning.

ECG Signal Quality Assessment and Arrhythmia Detection

• Developed a deep learning pipeline to assess ECG signal quality and detect arrhythmias, trained on 500 hours of ECG data
• Employed time and frequency domain features, along with key morphological markers (e.g., RR intervals, QRS width, P-wave variations)
• Applied signal preprocessing (bandpass filtering, baseline wander removal, and powerline interference reduction) to enhance data quality
• Utilized an XGBoost-based classifier to combine diverse features and signal-derived metrics for arrhythmia detection (e.g., AFib, VT)
• Validated the system against clinical reference annotations to ensure reliability

BCG Signal Processibility Classification for Vitals Estimation

• Developed an ML model to classify processable vs
• Non-processable BCG signals before vital sign estimation
• Identified and filtered out non-human artifacts such as sensor movement, environmental vibrations, and noise to ensure reliable signal quality
• Trained in controlled and uncontrolled BCG datasets, enhancing robustness for real-world applications
• Improved accuracy of heart rate and respiration estimation by ensuring that only high-quality signals are processed