Parkinson’s Disease (PD) is a progressive neurodegenerative disorder that significantly impairs motor and non-

motor functions. Early detection is critical for timely intervention, yet conventional diagnostic methods remain

limited, particularly in resource-constrained settings. This study presents a dual approach for Parkinson’s

Disease detection: a traditional non-AI clinical evaluation framework and a novel deep learning-based model

named NeuroParkNet. The clinical model relies on structured symptom evaluation, drawing tests, voice

recordings, and gait observations without the use of artificial intelligence, offering a cost-effective solution for

rural and underserved regions. Complementing this, the NeuroParkNet deep learning model processes spiral

drawings, Mel spectrograms from voice samples, and gait accelerometer data using a tri-stream architecture

composed of ResNet-18, Conv2D-BiLSTM, and Conv1D-GRU modules. Trained on a fabricated multimodal

Parkinson's Disease (PD) is a progressive neurodegenerative disorder that primarily affects movement. It is

considered the second most common neurodegenerative disease after Alzheimer's disease. In India, the burden of

neurological disorders, including Parkinson's, is steadily increasing due to rising life expectancy, lifestyle

changes, and lack of awareness. The disease, named after English doctor James Parkinson, who first described it

in 1817, affects the basal ganglia region of the brain, particularly causing the depletion of dopamine-producing

neurons in the substantia nigra [1]. The result is a constellation of motor symptoms like resting tremors,

bradykinesia (slowness of movement), rigidity, and postural instability. However, non-motor symptoms such as

sleep disturbances, mood disorders, and cognitive decline are equally significant and often overlooked [2].

In Indian clinical settings, the diagnosis of PD is primarily based on medical history, physical examination, and

India has an ageing population, and the incidence of PD is expected to grow significantly in the coming years.

The estimated prevalence is 70–328 per 100,000 in India, with higher rates in older age groups [4]. However, the

real number could be much higher due to underdiagnosis and misdiagnosis, especially in underserved

communities. Societal stigma, lack of awareness, and general neglect of geriatric health are key contributing

factors [5]. Moreover, in many Indian families, the early signs of PD—such as hand tremors or slow movements—

are dismissed as normal ageing rather than symptoms of a disease that requires medical intervention [6]. Detection

simple recording tools and basic software, making them viable in low-resource environments [9]. In fact, some

Indian neurologists use standard mobile audio recordings to assess vocal biomarkers like jitter, shimmer, and

pitch variation, which are helpful in differentiating PD from other disorders [10].

India’s medical infrastructure often lacks high-end imaging technologies like Positron Emission Tomography

(PET) or DaT SCAN, which are used in Western nations for PD diagnosis. In contrast, clinical diagnosis in India

is frequently supported by Magnetic Resonance Imaging (MRI) or Computed Tomography (CT) to rule out other

structural brain abnormalities. While MRI cannot confirm PD, it is used to exclude stroke, tumour, or

hydrocephalus, which might present with similar symptoms [11]. Blood and cerebrospinal fluid tests are rarely

used in routine Indian practice, mostly due to lack of established biomarkers and cost constraints. However,

posture, reduced arm swing, or short shuffling steps—is commonly used by clinicians, physiotherapists, and

caregivers to track disease progression [14]. India also benefits from its traditional and complementary medicine

systems. Some Ayurvedic practitioners and neurologists collaborate to assess pulse patterns and muscular rigidity,

supplementing modern diagnostic methods with traditional knowledge. While these approaches are not validated

globally, they do provide culturally accepted alternatives and support early detection when integrated properly

[15].

Parkinson's Disease (PD) is a chronic and progressive neurodegenerative disorder affecting motor and non-motor

functions. Early detection is vital for effective treatment and improving quality of life. This section reviews

various research papers focused on PD detection approaches, excluding redundant AI-centric perspectives already

indicating significant differences in motor control. This method provides a non-invasive and easy-to-administer

technique for clinical application. Jiao Meng et al. [17] proposed a model for detecting Parkinsonian tremors

using Discrete Wavelet Transform and Singular Value Decomposition. Their focus on tremor characteristics

reinforces the potential of physical signal analysis without relying entirely on AI systems. Tremor data serves as

a reliable clinical marker, particularly when enhanced through data augmentation. Yuzhe Yang et al. [18] explored

nocturnal breathing signals to detect PD progression. Although their model is AI-driven, the core insight—

correlating disease severity with sleep-related physiological data—opens new avenues for traditional sensor-

based monitoring that can be adapted in wearable technologies without advanced AI. Bahar et al. [19] developed

a method combining non-template drawing activities and principal component analysis to differentiate PD

Gait abnormalities are hallmark symptoms of PD and have been widely studied. Jadhwani and Harjpal [20]

reviewed gait patterns and proposed an automated approach for evaluating freezing of gait using wearable sensors.

While AI was used for final classification, the captured motion patterns and gait variables (stride length, cadence)

diagnosis and treatment adherence if user-friendly and transparent.

Voice alteration is one of the earliest PD symptoms. Shen et al. [23] investigated vocal biomarkers such as jitter

and shimmer to predict PD. Their approach, although supported by neural networks, identifies specific acoustic

features that can be captured and analyzed by simpler tools like signal analyzers in non-AI contexts. Roy et al.

[24] confirmed similar findings in a broader review of voice, gait, and EEG-based modalities. Their literature

synthesis underscores voice impairment as a universal marker, advocating for integration into basic health

MRI, PET, and EEG are traditional methods in neurological diagnostics. Shreya Reddy et al. [26] employed MRI

scans to differentiate PD patients. Although their work integrates AI, the anatomical differences observed in PD-

Apart from technical methods, diagnosis adoption depends on sociocultural factors. AlSafran et al. [29] evaluated

the feasibility of AI-driven PD diagnosis in the U.S. market. While centered on strategic implementation, they

identified essential diagnostic requirements such as symptom quantification, reliability, and clinician usability—

features applicable to manual or semi-automated systems as well. Dhanalakshmi et al. [30] conducted a systematic

review of emergent PD detection systems including handwriting, gait, and brain imaging modalities. Their work

critically analyzed the performance and limitations of each system, concluding that multimodal systems (not

necessarily AI-powered) yield better detection accuracy. Such findings validate combining basic diagnostic tools

(voice, hand movement, and EEG) in standard PD screenings.

early-stage Parkinson’s Disease using multimodal biomedical data. The model is designed to process spiral

drawing images, voice recordings, and gait sensor data simultaneously through parallel convolutional and

recurrent sub-modules. The objective of Neuro Park Net is to learn both spatial and temporal characteristics of

publicly available data and simulated patient records to ensure sufficient variability and clinical realism.

tablet inputs, while voice data was simulated using acoustic modeling based on vocal tremor and dysphonia

characteristics. The gait data was modeled on realistic walking patterns observed in PD patients, including step

asymmetry and stride irregularity.

To ensure consistency, all spiral images were normalized by converting them to grayscale, resizing them to

224×224 pixels, and applying histogram equalization. Voice clips were down sampled to 16 kHz mono-channel

WAV format and then converted into Mel spectrograms using a window size of 25 ms and a hop length of 10 ms,

resulting in fixed-size 128×128 spectrogram matrices. Gait data was segmented into fixed-length sequences of

300 time steps with three channels (X, Y, Z axes), and standard normalization was applied across the dataset to

align amplitude scales. All data modalities were synchronized at the patient level to maintain record integrity

during training.

also 128-dimensional.

The gait sub-network is a time-distributed 1D convolutional network followed by a Gated Recurrent Unit (GRU)

layer. This allows the model to capture temporal sequences of lower limb movement, including freezing episodes

and gait speed fluctuations. After temporal aggregation, a 128-dimensional vector is extracted.

The outputs of the three streams are concatenated into a 384-dimensional feature vector, which is passed through

a dropout layer to prevent overfitting. This joint representation is then processed through two fully connected

developed and executed using TensorFlow 2.0 on an NVIDIA RTX 3090 GPU with 24GB memory. Stratified 5-

fold cross-validation is applied to ensure generalization, and data augmentation is used on the spiral and voice

inputs to improve robustness.

and gait sensor time series. Performance was measured using standard classification metrics such as Accuracy,

Precision, Recall, and F1-Score.

and a Late Fusion Multimodal CNN baseline. The Neuro Park Net model achieves the highest accuracy at 96.8%,

outperforming the closest competitor—Late Fusion CNN—by 4.2%. In addition, Neuro Park Net shows

consistently higher values across all evaluation metrics, reinforcing its ability to generalise better across diverse

patient data.

Table 1: Performance Comparison of Parkinson’s Disease Detection Models