A Deep Learning Approach to Predict Knob Turning Activity with Varying Torque and Arm Angles Using Force Myography (under review)

This study investigates the use of force myography (FMG) to predict knob-turning activity with varying torque values and arm angles. Knob-turning is a complex task involving precise motor control and finger coordination. Despite its importance in daily activities, it has received limited attention. The study involved twenty participants who performed knob-turning activities using 3D-printed knobs with three torque values and four elbow angles. FMG data were collected during these activities. The study utilized a Long Short-Term Memory (LSTM) classification approach to classify the knob activities. The results show that FMG is an effective method for predicting knob-turning activity, torque applied, and arm angle. The LSTM-based algorithm achieved high accuracy rates: 98.21% for knob-turning activity, 95.40% for torques, and 94.96% for arm angles. The presence of subcutaneous fat did not significantly affect the classification accuracy. This shows that FMG can accurately predict knob-turning and have potential in real-time applications.

Freezing of gait (FoG) is a widely observed movement disorder in patients with Parkinson’s disease (PD). Its prediction is crucial for effectively giving the cue to avoid FoG occurrence. However, present methods of prediction of FoG are inaccurate for large but practical prediction horizons (PHs). Therefore, this work presents a comprehensive analysis of the electroencephalography (EEG) and inertial measurement units (IMUs) to predict FoG advance in time. An ensemble model consisting of two neural networks (NNs), EEGFoGNet and IMUFoGNet, was developed and tested at different PHs and ensemble weights. Moreover, the model is tested for two practical scenarios: clinical or research applications and personal uses. For clinical or research applications, stratified fivefold cross-validation was used. For personal uses, a transfer learning technique was used for learning user-specific FoG-related features. The model obtained the best accuracy of 92.1% at 1-s PH and the least accuracy of 86.2% at 5-s PH. The presented results are encouraging and show the proposed model’s clinical applicability. This study will also help practitioners in comparing the efficacy of different cueing methods.

Effect of Bhramari Pranayama on Brain: EEG studies [Study 1] [Study 2]

The findings of this joint study with All India Institute of Ayurveda New Delhi suggest that Bhramari pranayama, when performed with shorter durations of humming, has a positive effect on brain wave activity. This study explores brain waves for scientific validation of Bhramari Pranayam benefits. Moreover, the results indicate the potential development of a wearable sound recording system in the future as a feedback mechanism, providing biofeedback to users to maintain a consistent humming duration during the practice.

The study developed a novel, low-cost, and wearable instrumented nasal temperature sensing (NTS) device for measuring various breathing parameters. It showed promising applications in hyperventilation, Kussmaul breathing, sleep apnea, shortness of breath in COVID-19, bradypnea, sinus arrhythmia, and breathing pranayama. The device was tested during controlled breathing exercises in different positions and ambient temperature settings, as well as during standing and walking at various speeds. Comparisons with a commercially available respiration belt (RB) (gold standard) demonstrated that the NTS device is accurate for real-time applications, with a mean absolute error (MAE) ranging from 0.09 to 0.17 breaths per minute (bpm). Additionally, the device showed potential advantages of affordability, aesthetics, and robustness against motion artifacts.

Gait assessment scores are used for quantifying the abnormalities in the gait. Evaluation of the performance of these scores is a must for their clinical acceptance. However, current methods of assessing the performance of the gait assessment scores for clinically relevant gait abnormalities are prone to error. For example, values of intra-observer reliability, inter-observer reliability and sensitivity calculated for a gait assessment score change with the population of patients and observers. Therefore, there is a need for a methodology for replicating musculoskeletal deformations such as contracture in healthy individuals for objectively evaluating the performance of gait assessment scores with variable severity of musculoskeletal deformations. In this study, a series of dynamic musculoskeletal simulations are performed to simulate and verify a mathematical model of a passive exoskeleton for simulating contractures.

Affordable long-term recording of gait kinematics may accelerate research in sports, clinical, and personal biomechanics. Due to the limitations of the present methods, long-term motion capture during walking ensuring the user’s comfort has not yet been achieved. Therefore, a neural network-based model (foot2hip) for predicting ankle, knee, and hip joint angle profiles using foot kinematics and kinetics is developed in this study. Foot2hip consists of three convolution, two max-pooling, two LSTM and three dense layers. Indigenously developed insole and outsole were used to measure the kinetics and kinematics of the foot, respectively. Seven healthy participants were recruited to follow an experimental protocol consisting of six walking conditions namely, slow, medium, fast walking speed, rearfoot, flatfoot and forefoot landing pattern. When tested for leave-one-out and nested cross-validation, foot2hip obtained an excellent prediction performance. The results are encouraging and show foot2hip’s applicability in accurately predicting lower limb kinematics with minimal wearables.

In this work, a force myography-based step length classification model which can predict long and short steps before their completion is presented. Healthy participants walked over a surface marked with long and short steps while wearing a force myography system over their left thigh and a force-sensitive left insole. Three machine learning models were trained using the processed force myography signal to classify long and short steps. The machine learning model trained by the entire stride signal presented the highest F1-score of 86.64 % proving that the force myography signal of the thigh is a potential input signal for automated step length control in powered prostheses and assistive devices. Pdf file will be added soon.

In this work, we propose a novel center of pressure (CoP) based vibrotactile feedback step length training system to alter step length. Anterior/posterior location of the CoP at heel strike compared to the baseline was mapped to the anteriorly/posteriorly located tactors on the waist band. Pdf file will be added soon.

This paper presents a neural network model having three modules for classifying walking patterns into nine known knee joint abnormalities in the sagittal plane. The network obtained the best classification accuracy of 98%, the precision of 0.93, recall of 0.95 and f1-score of 0.95. These results suggest that a neural network-based method can be used as a gait assessment tool for known gait abnormalities.

Proposed new foot kinematics measuring device for gait assessment of CP children. The developed system was validated against the reference marker-based motion capture system from Noraxon, Scottsdale, Arizona, USA. The data from eight able-bodied participants were acquired simultaneously from both the systems (developed and the reference system) at three different walking speeds and three foot landing strategies. 

Automated-Gait Assessment Score (A-GAS) is the first objective, comprehensive gait assessment methodology developed for CP children. A-GAS report has three components corresponding to different levels of gait analysis. The first component depicts the global abnormality index of a limb, the second component depicts the abnormality index of each joint, and the third component depicts the abnormality of each joint at multiple gait cycle instances. For better clinical acceptance, A-GAS is developed such that the methodology behind the algorithm of A-GAS resembles the thinking process of a clinician while comparing an abnormal gait with a normal gait. Therefore, the parameters of A-GAS are easy to understand and manipulate.

Response of the body during Yoga

Developed a new device for recording respiratory activities during Yoga (Indian design patent). Validation of the device is available online [Link]

Reference joint angle profiles are used by most of the advanced prostheses and exoskeletons. Recording these profiles at various working conditions of the exoskeleton can be a tedious task. To make this process smooth, this study proposes a neural network that can generate knee joint angle profiles for various conditions using minimal input. The proposed model obtained a mean absolute error of 2.66℃ ± 1.00℃ from 10-fold cross-validation.

Exploring how Depression affects human brain using artificial intelligent systems. The project tried to understand the difference in activities of the brain of depressed subjects and healthy subject using electroencephalography (EEG). The study is completed and the developed system can detect depression level in a subject with an accuracy of 99.37% on the bases of his/her EEG recordings.

Linguistics is the best way of understanding the workings of the human brain. The project aims at detecting text regions in an image by using a convolutional autoencoder. Here, the fundamental features present in a written piece are extracted by a neural network to separate the text segment from the background in an image. We used a light weighted and simple framework to achieve the goal.

Developed a classifier using a convolutional autoencoder and k-means clustering algorithm for detection eye state using EEG data of 14 channels of 4 seconds with an accuracy of 100%. The proposed method can be used for reducing the size of EEG data with a compression ratio of 96.875% with insignificant lose in data.

Multisensor Integrated car fire alarm

Developed a cost-efficient car fire detection system using real-time processing and integration of data from various sensors.

Hand gesture-controlled home automation

Developed a fast and robust hand gesture controlled home automation framework. Used self-fabricated microcontroller module for smartly controlling devices by detecting the intention of the users via their hand gesture.

Developed a computational tool for fast fabrication of mechanical parts using CNC machine from engineering drawing (ED). The software extracts significant pieces of information from the given ED, convert it into vector format using a novel algorithm, and generate the path design in the form of g-code. This tool also allows user to store the file in compressed form (c/r: 98%). 

2-D CNC machine

Developed a GUI enabled CNC machine which can be used for 2-D applications. Keywords: Image processing, inverse kinematics, trajectory algorithms

Developed and deployed the first working product for automatically controlling advertisement screens across the cities in India. It includes the development of a Linux-based software "Consilium" for remote advertising screens.

Developed a soft social robot for children having Down Syndrome. It was the first cost-effective, fully functional and practical intervention tool for the all-round development of children having Down Syndrome.

Worked as captain of the institute's technical team for two years. Worked in the mechanical, electrical and coding department. Keywords: CAD modelling, 3-D printing, fabrication, sensors, processing boards, and image processing.

An augmented reality-based product was created that could replace the dress changing room of the shopping mall.

Short and smooth path planning for a mobile robot such that it would consume lesser time and fuel. Developed a computer program by using numerical methods like interpolation, linear equation solving, differentiation, and splines.