Machine learning for biomedical data analysis


Biomedical device & algorithm development for quantitative assessment of neurological disorders.

Cerebellar Ataxia & Friedreich Ataxia

    • Speech
    • Upper Limb
    • Lower Limb
    • Balance
    • Gait

Development of human motion capture and analytic systems for medical and sports applications.

Biomedical device & algorithm development for quantitative assessment of the neurological disorders – Cerebellar Ataxia & Friedreich Ataxia in the following domains:


This research focuses on the quantitative assessment of speech motor disorders in neurological disorders. It is focused on designing a self-diagnosis tool to analyze a patient’s voice, identify symptoms, calculate the severity of the speech disorder, and predict the progression.

Globally, 90% of the population with Neurological Disorders (ND) have speech problems with symptoms worsening over time. There are no laboratory biomarkers that can detect ND, and brain imaging scans do not allow for a definitive diagnosis. Through her research, Bipasha employed acoustic signal processing techniques to identify a nexus between the presence of certain speech impediments including a slurry voice, and the existence of a neurological disorder. Using statistical and machine learning techniques, she has developed an algorithm that can extract unique characteristics from a patient’s voice as they speak to their phone, and this produces a score.  This score indicates the presence of an underlying neurological condition and the extent of its severity.

Such a technique to analyze speech and vocal patterns would be effective for diagnosing ND, and possibly at earlier stages than is now possible. With early intervention, therapeutic measures can be adopted to control disease progression. Such a self-diagnosis system will not only empower patients to have greater control of their healthcare decisions but also aid the doctor to make informed clinical decisions.


Cerebellar damage can often result in disabilities affecting the peripheral regions of the body. These include poor and inaccurate coordination, tremors, and irregular movements.

Dysdiadochokinesia (DDK)

This work uses feature extraction techniques to identify features within the kinematic information gained from kinematic measures, that predict the presence and severity of ataxia as judged by the clinician. Limb movements associated with these tests are measured using BioKinTM for real time data capture and transmission. The subject is instructed to pronate their hand, so that palm side faces downwards to rest on the palm of the other hand. The subject is also instructed to pronate and supinate between these two positions with maximum speed, and the rate of alternation is extracted from the BioKinTM attached to the wrist. Kinematic parameters such as acceleration, velocity and angle are considered in both time and frequency domain in three orthogonal axes to obtain relevant disability related information. The sensory information captured from the test is analyzed to uncover certain features intrinsically linked to the underlying disability as well as information describing the extent of the disability. Machine learning approaches were used to enhance the classification and severity assessment.

Finger Nose Test (FNT)

The subject was instructed to use their index finger to first touch the clinician’s finger and then their own nose and repeat this task for approximately 15 s. The test is performed for left and right limb consecutively. The clinician holds their finger at a stationary position during the task. The BioKinTM unit is attached to the subject’s dorsal surface of the hand. The feature extraction technique and methodology is as same as DDK.

Ballistic tracking (BAL)

In order to establish inter rate concordance with enhanced reliability and effectiveness in the assessment of upper limb function, a novel automated system employing Microsoft KinectTM was developed to capture the motion of the CA patient’s finger for objective assessment. This essentially mimics the commonly used finger tracking task clinically assessed through subjective observation. Thus generate an automated system eradicating clinician’s subjectiveness.





Heel-shin Test (HST)

The IMU sensor was strapped to the ankle as depicted in Fig. 1b. The subject is seated comfortably on a chair with their eyes opened and the right heel is placed on their left shin below the knee and slid down along the leg towards the heel. This motion was repeated on each side for ten cycles. The SARA subscore 8 is used for the clinical assessment of peripheral dysfunction via the Heel-shin test quantified by a numerical value from 0 to 4. The Heel-shin task has an advantage in assessing non-ambulant individuals.


Romberg (ROM)

Utilization of motion sensor in quantitative assessment Cerebellar Ataxia with the balance Romberg’s test.


The human ability to walk normally relies on both trunk and limb coordination. The cerebellum participates in this coordination through rhythmic inputs to the locomotor brainstem centres for supra-spinal locomotor control and an integrated sensory input to achieve coordinated limb movements. Gait ataxia and truncal instability are common manifestations of injury or dysfunction of the midline cerebellum or vermis. An ataxic gait is characterised by a wide-base (increased distance between the feet), an unstable walking path, and high variability of cadence. This marked variability in gait leads to an increased risk of falls. The gait-related ataxia studies at our lab help clinicians in the diagnosis and severity estimation processes.