Module - 5

WEARABLES TO THINKABLES
Remote Health Monitoring using Wearable Sensors
Definition: Remote health monitoring involves collecting information about a person's bodily functions
without the need for physical presence.
Technologies Used:
• Video or phone calls for remote consultations.
• Mobile apps or websites for sharing images and data with medical experts.
Wearable sensors for continuous data collection and logging on mobile devices or in the cloud.
Types of Wearable Sensors:
1. Consumer-grade wearables: Commonly used by individuals for fitness and basic health
monitoring.
2. Clinical-grade wearables: Used in medical settings for more precise monitoring.

USE CASE – 1
Monitoring Patients with Parkinson’s Disease
Symptoms Monitored: Slowed movements (bradykinesia), rigidity, tremor, frequent falls.
Sensors Used: Motion sensors or inertial measuring units (IMUs) such as accelerometers, gyroscopes, and
magnetometers. These sensors can be attached to clothing or body parts.
Example System: Parkinson’s KinetiGraph (PKG) by Global Kinetics.
Components: Wrist-worn device, algorithms for analyzing movement patterns, and reporting tools.
Process: Data is collected over several days, analyzed to score movement patterns, and results are provided
to clinicians in reports.

USE CASE – 2
Monitoring patients with type I diabetes mellitus: Glucose monitors
Definition: Autoimmune destruction of pancreatic beta cells, leading to no insulin production.
Management:
- Lifestyle and dietary changes.
- Monitoring blood glucose and administering insulin:
- Too much insulin: Hypoglycemia (drowsiness, confusion, coma, death).
-Too little insulin: Hyperglycemia (risk of diabetic ketoacidosis, long-term complications like
cardiovascular disease, renal failure, neuropathy).
DIY Systems:
- Examples: OpenAPS, AndroidAPS, Loop.
- Open-source, patient-used, but unregulated and not clinically validated.

USE CASE – 3
Monitoring Rehabilitation Exercise Adherence
Objective: Ensure patients follow prescribed rehabilitation exercises.
Technology: Wearable sensors track movements and provide feedback.
Benefits: Improves recovery rates, provides real-time data to therapists, and increases patient
accountability.
Challenges: Ensuring accuracy of data, patient compliance, and integration with healthcare systems.

USE CASE – 4
Continuous Monitoring of Individuals Using Consumer-Grade Wearables
Objective: Track health metrics like heart rate, sleep patterns, and activity levels.
Technology: Smartwatches and fitness trackers.
Benefits: Provides continuous health monitoring, early detection of health issues, and promotes healthy
lifestyle choices.
Challenges: Data privacy concerns, accuracy of consumer-grade devices, and user engagement.

USE CASE – 5
The Use of Consumer-Grade Devices for Cognitive Impairment Monitoring
Objective: Monitor cognitive functions in individuals with impairments.
Technology: Wearables with cognitive assessment tools and apps.
Benefits: Early detection of cognitive decline, continuous monitoring, and support for caregivers.
Challenges: Ensuring data accuracy, user compliance, and integrating data with medical records.

USE CASE – 6
eHealth and Medical IoT Development Platforms
Objective: Develop platforms for integrating various health monitoring devices.
Technology: IoT devices, cloud computing, and data analytics.
Benefits: Centralized health data management, improved patient care, and enhanced data analytics.
Challenges: Ensuring interoperability, data security, and managing large volumes of data.

• Sport performance monitoring

Athlete performance can be closely monitored using sensors for electromyogram, electrocardiogram,
pulse oximetry, blood pressure, airflow, and body weight scale.
• Heart monitoring
Patients suffering from hypertension or arrhythmias can be monitored using this kit comprising sensors
for pulse oximetry, blood pressure, electrocardiogram with patches, and an
alert alarm button
• Pregnancy monitoring
Developers can use this kit to monitor and create insights throughout pregnancy. There are
sensors for monitoring blood pressure, body scale, and a glucometer with strips.

• Respiratory monitoring
This kit is ideal for developers wishing to address persons suffering from respiratory disorders such as
asthma or chronic obstructive pulmonary disease. It comprises a spirometer
with mouthpieces, pulse oximetry, and airflow sensors.
• Sleep apnea and snore monitoring
Specially designed to monitor patients with sleep disorders that can produce apnea (suspension of
breathing), this kit comprises pulse oximetry, body position, airflow, snoring, and
blood pressure sensors
• Elderly person monitoring
The monitoring of elderly persons for blood pressure, blood glucose, and oxygen can be
tracked. It is also suitable for the monitoring of those suffering Alzheimer’s or others affected
by dementia and can monitor a patient’s position. For alerting support persons an
emergency alarm button is included.
• Endocrine disorders monitoring
Patients with endocrine disorders, and who may require keeping track of their weight and
muscle mass variations as well as their blood sugar levels can benefit from this monitoring kit.
Sensors comprise a glucometer with strips and a body weight scale

Platform and Sensor Device Integration

Connected sensors collect data and store it on the device before securely sending it via WiFi or Bluetooth
to a private cloud account. A developer API allows access to this data for visualization and integration
into third-party systems. The My Signals device supports pairing with a variety of Bluetooth sensors and
offers a development platform compatible with Arduino. This setup enables custom hardware integration
and on-device data analysis. For instance, it can be used to monitor sleep apnea, optimize athlete workouts,
or analyze sleep and nutrition data for personalized schedules.

AI Enabled Sensors and its Challenges

Definition: Artificial Intelligence (AI) involves technology that learns from data, reasons based on
insights from data analysis, and interacts with humans in a human-like manner1. AI models adjust
themselves during training to improve performance, aiming to match or exceed a defined level for real-
life deployment2. This process involves discarding unfavorable changes and preserving beneficial ones,
ultimately enabling AI to analyze new, unseen data effectively.

AI at the Edge: From Wearables to THINKables

AI at the edge refers to the implementation of artificial intelligence (AI) algorithms directly on wearable
and edge devices, rather than relying on cloud computing. This approach enhances power efficiency and
real-time processing by minimizing data transmission. It involves using dedicated hardware like
application-specific integrated circuits (ASICs) and field-programmable gate arrays (FPGAs) to optimize
performance and reduce power consumption. The emergence of platforms like TensorFlow Lite has made
it feasible to deploy deep learning models on these devices, enabling advanced functionalities while
maintaining low power usage and flexibility.

Use of THINKables in health monitoring:

The use of THINKables in health monitoring discusses the integration of wearable technology into
healthcare for continuous and unobtrusive monitoring of patients. THINKables, which are advanced
wearable sensors, can track vital signs, detect potential health issues early, and provide real-time data to
healthcare providers. This technology aims to enhance patient care by enabling preventive measures,
reducing the need for frequent hospital visits, and improving overall health outcomes through constant
monitoring and timely interventions.

Different Methods used in AI:

It covers a variety of approaches and techniques that enable artificial intelligence systems to perform tasks
that typically require human intelligence. These methods include machine learning, where algorithms
learn from data to make predictions or decisions without being explicitly programmed. Within machine
learning, there are subcategories like supervised learning, where models are trained on labeled data, and
unsupervised learning, which involves finding patterns in unlabeled data. Reinforcement learning is
another method where agents learn to make decisions by receiving rewards or penalties based on their
actions.
Another significant method is deep learning, a subset of machine learning that uses neural networks with
many layers to analyze complex data. Natural language processing (NLP) enables AI to understand and
generate human language, facilitating applications like chatbots and language translation. Computer
vision allows AI to interpret and process visual information from the world, enabling tasks like image
recognition and autonomous driving. These methods collectively contribute to the advancement of AI,
making it capable of performing a wide range of tasks across various domains.

Fig: THNIKable architecture

 Fig: Raspberry-Pi coupled to intel Movidius Neural Compute stick and camera

USE CASE 1:
“The digital epilepsy diary” discusses the development and benefits of a digital diary for epilepsy
patients. This diary helps patients and healthcare providers track seizures, medication, and other relevant
data. By using wearable sensors and mobile technology, the diary provides real-time monitoring and data
collection, which can improve the management and treatment of epilepsy.
The digital epilepsy diary also offers features such as reminders for medication, alerts for unusual patterns,
and the ability to share data with healthcare professionals. This technology aims to enhance patient
engagement, improve adherence to treatment plans, and ultimately lead to better health outcomes for
individuals with epilepsy

USE CASE 2:
Smart sensors for exercise tracking:
Consumer-grade wearable sensors like the Apple Watch and Fitbit can track calories burned and heart
rate during exercise. Machine learning enhances these devices by providing feedback on exercise
techniques, crucial for preventing injuries caused by poor form. Kowsar et al. developed a smart sensor
prototype that learns an athlete’s correct posture and uses anomaly detection to identify deviations due
to factors like muscle fatigue. When a deviation is detected, an alarm alerts the athlete to correct their
posture or stop the exercise.
USE CASE 3:
AliveCor for detection of atrial fibrillation:
Atrial fibrillation (AF) affects over 5 million Americans, with nearly 700,000 undiagnosed cases. AF can
cause symptoms like palpitations or dizziness and increase the risk of stroke. Traditional
electrocardiograms (ECGs) face barriers such as cost, interpretation, and accessibility, making it
challenging to diagnose cardiac conditions effectively. Intermittent arrhythmias, like paroxysmal atrial
fibrillation, may be missed by one-off monitoring, even if the ECG is obtained over an extended period.
Wearable devices, such as the AliveCor, offer continuous ECG monitoring or can be activated to perform
an ECG immediately if symptoms develop, eliminating the need for a medical facility visit. The AliveCor
system includes a sensor, an iPhone or Apple Watch, and an application. The sensor measures the electric
potential between the user’s fingers, transmitting data to the iPhone microphone via an ultrasound signal.
This data is digitized and converted into a real-time ECG trace, which is noise-reduced and processed to
detect AF.
The AliveCor system exports ECGs to a PDF or email and uploads results to a secure server for review
by a physician or a remote cardiologist. It integrates with Apple’s Health App to share key health
parameters. The device is suitable for users of all ages and has received FDA and TGA approval, as well
as the CE mark. Trials have shown its efficacy and cost-effectiveness in screening asymptomatic patients
for atrial fibrillation.

USE CASE 4:
Nail sensor for detecting and monitoring Parkinson’s disease progression:
In December 2018, IBM Research introduced a groundbreaking prototype sensor device that monitors
human health via the fingernail. This device integrates strain gauges and an accelerometer, allowing it to
detect various signals from the bending of the fingernail as the user performs everyday tasks. One key
insight derived from this is an indicator of a person’s grip strength. The project initially aimed to monitor
medication state and disease progression in individuals with Parkinson’s disease, who often have fragile
skin that complicates the use of direct skin-based sensors.

The fingernail sensor device collects data and sends it wirelessly to a smartwatch, which processes the
data using machine learning models. These AI models can identify finger movements, physical activities
involving the hand, and characters written with a stylus. Additionally, the device enables monitoring of
Parkinson’s disease symptoms such as bradykinesia, tremor, and dyskinesia, providing a non-invasive and
reliable method for tracking disease progression.
Challenges of AI-enabled sensors in health

• Integrating AI techniques into wearable systems does not come without hurdles: medical data is
highly sensitive causing some of the biggest challenges to be centered around data security, privacy,
and access as well as around building ethical AI models. It is not enough for AI algorithms to produce
the desired detection or prediction results – they also need to be transparent, explainable, and fair to
gain regulatory approval and the trust of users.
• These features are often referred to as the “ethics of AI,” a focus of current research which is just as
important as developing the algorithms themselves. But these challenges are not insurmountable.
Regulatory bodies, biomedical, pharma, bio-tech, AI-companies, start-ups, as well as medical and
clinical research organizations have all started to jointly tackle these challenges and to develop
frameworks for exploring a variety of AI and Blockchain technologies for patient monitoring.
• What we see at this point are mostly early-stage feasibility pilot studies demonstrating the high
potential of AI techniques for improved patient monitoring. Rigorous research and development are
necessary to ensure the viability of these innovations before THINKables can be fully integrated into
clinical workflows.
• Any breach of a research protocol or the premature setting of unreasonable expectations may lead to
an undermining of trust and ultimately the success of THINKables in the healthcare sector.

The core aspects of the ethics of artificial intelligence:

1.Accountabilty 2.value alignment 3. Explainability 4. Fairness 5. User Data right

Data analytics for wearable IoT-based telemedicine

Modern wearable technology demands a special set of criteria to balance between device design and data
processing performance. It has been the trend for data analytics to be pushed down the pipeline from the
cloud to an edge or wearable device. Along with improvements in processing power, network capabilities
are also expanding such as with the case of 5G, which is promised to improve speed and accuracy in
communicating with billions of devices. The data analytics pipeline is presented with an example of a
smart textile glove. The glove is integrated with an Inertial Measurement Unit (IMU) commonly found in
smart wearables, such as smartwatches and other wrist-worn devices, which are designed to track body
movements.
The need and demand of wearable technologies in the society
Wearable technologies, such as smartwatches and fitness trackers, have revolutionized how we monitor
and interact with our environment. These devices often include sensors that can track physical activity,
heart rate, sleep patterns, and more. For example, the Garmin Venu 2 Plus is a touchscreen sports watch
that offers comprehensive health monitoring features. Similarly, the Movano ring is designed to provide
health insights specifically for women.
The importance of sampling frequency, as you mentioned, is crucial in ensuring accurate data collection.
The Nyquist frequency, which states that the sampling rate should be at least twice the highest frequency
of the signal being measured, is fundamental in signal processing. This principle helps in capturing the
necessary information without losing any critical details.
Wearable technologies continue to evolve, offering more sophisticated and accurate ways to monitor our
health and activities. They provide valuable insights that can help individuals make informed decisions
about their lifestyle and health.

Smart Gloves
Unlike smartwatches that are designed for the measurements of daily life activities such as step counts,
calorie count, exercise, sedentary periods, and sleep, the smart glove is a specialized wearable technology
aimed at providing clinically relevant measurements of symptoms for Parkinson’s disease. Although the
gloves can be used for other applications in the healthcare industry, this chapter solely focuses on how the
gloves can measure the severity of Parkinson’s disease.

The Smart Glove design includes

• Custom Layered Cotton gloves designed with slips for quickly integrating various sensors and CPU
modules
• nRF51822 System on Chip (SoC) (BLE Nano Embedded Board)
32-bit ARM Cortex-M0 CPU
256 KB flash and 32KB RAM
2.4GHz radio supports Bluetooth Low Energy (BLE)
10-bit ADC @ 14.5 kHz
Two-Wire Master (for I2 C)
• Five flex sensors (one sensor for each finger)
• Inertial Measurement Unit (IMU) with 16-bit data output.
3 axis +/- 16 g linear acceleration full scale (accelerometer)
3 axis+/-16 gauss magnetic full scale (magnetometer)
3 axis +/- 2000 dps angular rate full scale (gyroscope)
• Android tablet with BLE capabilities to host companion applications for collecting data.
Smart Gloves Components and Design Considerations:
e-Textile glove design:
As shown in Figure 2, the flex sensors are placed on each finger using slips to hold each sensor in place.
The microprocessor, IMU, and recharging circuitry are integrated onto a single printed circuit board (PCB)
and enclosed within a 3D-printed box. The housing is adhered to the backside of the glove using Velcro
to allow for quick removal.

BLE nano embedded board:

BLE Nano is a Bluetooth 4.1 Low Energy (BLE) development board, which is one of the smallest ones.
Its core (Nordic Nrf51822) has an ARM Cortex-M0 System on a Chip (SoC) with BLE capability. It is
running at 16MHz and operates under both 1.8 V and 3.3 V so it is compatible with a lot of electronic
components.

Inertial measurement unit (IMU):

I 2 C communication is used to exchange data between the BLE Nano Embedded Board and the IMU
sensor (LSM9DS1 IMU IC). This consists of a three-axis magnetometer, three-axis gyroscope, and three-
axis accelerometer. The accelerometer and gyroscope in the glove are utilized to track orientation and
angular velocity, respectively. The gyroscope measures rotation across the 3-axis, whereas the
accelerometer measures movements in three dimensions, such as forward/backward, right/left, and
up/down. For the accelerometer, they both supply the raw measurement data as ax, ay, az, and for the
gyroscope, as gx, gy, gz. We can obtain more trustworthy movement data by combining those raw data
and translating them into orientation and angular velocity [36]. The idea of accelerometer and gyroscope
measurements is depicted in Figure 3.
Flex sensors:
A resistor network configuration is paired with each piezoresistive flex sensor to create a voltage divider,
which is sampled at 126 Hz by a 10-bit ADC. The flex sensors are designed to measure the angular
displacement at the finger knuckles. The resistance of the sensor increases when the finger bends,
producing an output that is comparable to the bend's radius.

Power limitations:
Wearable devices are limited by power constraints, and the required data quality for applications
influences power needs. ARM Cortex-M processors offer various capabilities, such as data processing,
memory pipelines, and timers. Key factors like instruction set architecture, clock speed, DSP extensions,
built-in caches, tightly coupled memory, and pipeline stages determine the processor’s workload. While
more complex on-board signal processing can enhance real-time response, it also increases cost and power
consumption. Security features like Trust Zone and Memory Protection Units are essential for medical
and military devices but less critical for consumer devices. Precision features, including Floating Point
Hardware and Sy stick Timers, improve accuracy. Coprocessor interfaces and integrated interrupt
controllers can reduce the main processor’s workload. Some boards offer multiprocessor solutions, with
one processor handling general tasks and another managing sensor fusion or network tasks. Processor
choice depends on the application, and offloading processing to the cloud can help save on-board power.

Transmission: A bottleneck in throughput

The firmware in the nRF51822 chip manages data from both the IMU and flex sensors in a smart glove.
This data is sent over Bluetooth Low Energy (BLE) to a connected device like a smartphone or tablet. An
Android app called "Smart Glove" receives this data and saves it into a CSV file for backup. These CSV
files are then sent via Wi-Fi to a server using the MQTT protocol for real-time monitoring and further
analysis.
Due to Bluetooth bandwidth restrictions, the sensor data, originally sampled at 126Hz, can only be
transmitted at a rate of 40Hz. This limitation causes some data packets from the flex and IMU sensors to
not reach the Android app, as shown in Figure 4.
Signal processing pipeline: From sensor signals to classifications
In the previous stage, sensor signals from the wearable device were transferred to a more powerful tablet.
Here, some signals were cleaned and labeled. The cleaned data is then sent to a computer server in its raw
form, where it undergoes preprocessing to reduce errors and noise. An overview of this signal-processing
pipeline is depicted in Figure 5, outlining each step conducted offline on the computer server.
However, each component of this pipeline will be explained separately to trace the raw signal through
various stages until it becomes a set of features ready for classification using Support Vector Machines
(SVMs) [38]. It's important to note that the initial raw signal is affected by data loss during wireless
transmission of data packets. As a result, although the signal was originally sampled at 126Hz, bandwidth
limitations mean that samples are recorded at approximately 40Hz.

Sample rates and quantization error

In order to recover the original signal from the lossy signal, we must understand the signal components
of interest to ensure that the pre-processing does not change the various features within the signal. The
IMU signals are sampled at the rate of 126Hz as signed 16-bit integer values on the hardware while the
wearer performs the hand flipping exercise; arms extended while rotating the hand palm down to palm up
and back repeatedly for 10 s.
 The speed of healthy adults guided our sampling rate while ensuring it satisfied the Nyquist theory. The
Nyquist theorem states that an analog signal can be digitized without aliasing error if and only if the
sampling rate is greater than or equal to twice the highest frequency component in a given signal.

where fs is the sampling frequency, which satisfies the Nyquist criterion and fmax is the highest frequency
in a given signal. The maximum angular velocity of human forearm pronation and supination was
estimated to be 300 dps. This corresponds to a 1.667Hz period for a flipping cycle, since our sampling
rate (125Hz) is much greater than 2Hz allowing us to comprehensively measure the hand flip. The
quantization error, which is dependent on the ADC resolution, is minimal due to the use of an IMU, which
uses a 16 Bit ADC measuring 65536 levels (216). The LSM9DS1 IMU used has a zero-g offset of 90mg
and angular zero rate offset of 30 degrees/s. The zero-g offset refers to the deviation of the measured
output accelerometer signal from the ideal output signal when no acceleration is present. Similarly, the
zero-rate level describes the gyroscope output signal when there is no angular rate present. These errors
can be countered by measuring the mean IMU values while the glove is in its default resting orientation
and compensating for the same while doing signal analysis.

Overcoming the transmission bottleneck with interpolation

The BLE 4.1 protocol's limited bandwidth, allowing a maximum transmission of 20 bytes per second as
discussed in Section 3.5, created a bottleneck in our system. Due to our application's lack of buffering,
not every sample point could be sent to the smartphone app, resulting in inevitable data packet loss. This
was a deliberate design decision, based on our confidence that we could still recover the signal even
without buffering.
To compensate for the lost data values, we used linear interpolation. This method estimates the missing
data points based on the known values. Since our application's Nyquist rate is 4Hz (twice the maximum
frequency), interpolation does not introduce aliasing errors. Figure 6 illustrates both the original signal
and the reconstructed signal using interpolation.
Sensor Fusion and Basics: Sensor fusion combines data from different sensors to enhance accuracy. In
our case, we use accelerometers and gyroscopes to determine orientation. Accelerometers sense changes
in direction due to gravity, while gyroscopes detect orientation through angular velocity changes.
However, gyroscopes tend to drift over time. By adding accelerometers, we minimize gyroscope bias and
improve orientation readings.
The Kalman Filter: The Kalman Filter is a powerful technique for sensor fusion. It optimizes estimates by
adjusting an averaging factor based on predicted uncertainty (covariance). Its recursive nature makes it
ideal for wearable devices with limited processing power. By combining accelerometer and gyroscope
data using the Kalman Filter, we achieve more accurate orientation estimates.
Analyzing IMU Data: To analyze Inertial Measurement Unit (IMU) data, we convert raw measurements
into angular units. Calibration corrects static errors, and we utilize accelerometer information to find Euler
angles—the system’s orientation. This process ensures reliable and accurate results for tracking and
orientation estimation.

Peak detection
Peak and valley detection is applied to pitch and roll signals from the extended Kalman filter. Using
MATLAB’s find peaks function, conditions like MinPeakDistance and MinPeakHeight are set to identify
peaks. Hardcoding these values isn’t ideal due to individual variances, so values are derived
programmatically by analyzing the signal. To avoid false peaks and valleys from idle data at the start and
end, a windowing operation using the Hanning window is applied, which dampens these sections.
For MinPeakHeight, the mean of the signal is used as the threshold for peak detection. MinPeakDistance
is determined by finding the maximum frequency through Fast Fourier Transform (FFT) and calculating
the step size by dividing the sampling frequency by this maximum frequency. This step size is then used
as MinPeakDistance during peak detection.
Valleys are detected by subtracting the signal from its maximum value before applying the peak detection
process. This method ensures accurate identification of peaks and valleys by considering the signal’s
characteristics and reducing the impact of irrelevant data.

Feature extraction
The feature extraction is performed on both the approximated pitch and roll signals obtained from the
extended Kalman filter (Figure 9). Features are measurable variables derived from the input data which
help describe it. The features, in this case, are obtained from the pitch and roll signals, features are chosen
by how well they vary between healthy participant data and PD participant data. The features chosen are
shown in Tables 3 and 4 indicates feature parameters on participants.

Human study and dataset description

We recruited five healthy participants and five patients with Parkinson’s Disease (PD), all aged between
60 and 80. Both groups performed the hand flipping task with their right hand first, followed by their left
hand, similar to a neurologist’s office procedure. Data was collected for 10 seconds per hand, resulting in
20 seconds of data per participant. The study received approval from the Institutional Review Board
(IRB).
Data was gathered using accelerometers and gyroscopes, excluding flex sensor data since fingers were
not involved in the task. Each glove, equipped with a 3-axis accelerometer and gyroscope, transmitted six
channels of data (12 bytes in total).
Classification example using SVM
For validation of the ML model, we use the technique of k-fold cross validation wherein:
1. Data is first shuffled and split into k fold
2. Each unique section is treated as a test set while a model is trained on remaining sections a. Validation
scores obtained on the chosen test set
3. Estimate the accuracy of the model by averaging accuracies for the model in each section
4. Summarize the skill of the model using the aggregate of model evaluation scores
In our study, we employed 5-fold cross-validation. We used 12 features extracted from the pitch and roll
signals as inputs to a Support Vector Machine (SVM) classifier. SVM is a type of classifier that aims to
find the best possible line or plane (hyperplane) to separate different classes of data points. To accomplish
this, the SVM identifies support vectors, which are data points crucial for determining the hyperplane's
position. These points are the ones nearest to the hyperplane and removing them would change the
hyperplane's location.
While the basic version of SVM works with linearly separable data, nonlinear separation between classes
can be performed by using various kernels. Kernels enable transforming data to a higher dimensional
space where the data is linearly separable enabling the use of SVM. This transformation enables a better
classification of data with nonlinearity. Our classification model applied a fine Gaussian kernel.
Figure 10 shows three confusion matrices: one for overall model performance (left), and two separate
matrices for left hand and right hand (center and right). Table 5 provides the accuracy results of the SVM
classifier. To assess the model's effectiveness, we use metrics like precision, recall, and accuracy. Precision
indicates the proportion of relevant results among the predicted ones, while recall denotes the fraction of
relevant results correctly identified by the model. Accuracy represents the overall percentage of correct
classifications in the dataset. These metrics help us gain a comprehensive view of how well the model
performs in its predictions.

Discussion
The proposed model effectively classifies hand flipping data as either healthy or with Parkinson’s Disease
(PD) based on a set of features. The left-hand classifier correctly identifies three subjects with PD and
four healthy subjects, but it incorrectly classifies one healthy subject as having PD and two PD subjects
as healthy. The right-hand classifier performs better, with only two false positives and no false negatives.
The scatter plots for the left and right hands show that healthy and PD features are more distinct in the
right hand, reflecting factors like user handedness. The case study has several limitations, including a
small dataset of five healthy participants and five PD patients, which is insufficient for robust machine
learning. The study did not consider the affected side for PD patients or the dominant side for all
participants, and only used the hand-flipping task. Future improvements could include better feature
engineering, ensemble models, and additional movement tasks. Infrared motion-sensing cameras offer
higher accuracy but have limitations like fixed data collection locations and the need for multiple cameras
to avoid occlusion.