This paper introduces TeleRehab AI, an intelligent mobile-based physiotherapy system designed to assist

patients in performing rehabilitation exercises independently. The system combines conversational AI with

real-time pose analysis to deliver personalized and interactive guidance. It integrates four major components:

a fine-tuned LLaMA 3.2 3B language model for physiotherapy consultation, a structured multi-layer message

processing pipeline, an on-device pose estimation module using Google ML Kit BlazePose, and a lightweight

exercise classification model deployed via TensorFlow Lite. The system evaluates squat performance using

biomechanical parameters such as knee angle, torso inclination, and movement depth. A Support Vector

Physiotherapy is a critical component in the recovery process for individuals affected by injuries, surgical

procedures, and long-term musculoskeletal disorders. Despite its importance, access to professional

physiotherapy services remains limited due to factors such as high treatment costs, uneven distribution of

healthcare professionals, and geographical constraints, particularly in developing regions [1], [15]. As a result,

many patients are required to perform rehabilitation exercises independently at home, often without adequate

supervision, which can lead to incorrect posture, reduced effectiveness, and increased risk of further injury

[10]. In addition, a significant number of patients fail to complete their prescribed exercise routines due to

lack of motivation and feedback, highlighting challenges in adherence and engagement [1].

These limitations emphasize key issues in current rehabilitation practices, including restricted availability of

enables users to perform exercises at home with real-time feedback. The system utilizes mobile-based

computer vision techniques along with conversational artificial intelligence to provide an accessible, cost-

effective, and engaging rehabilitation experience. By leveraging on-device pose estimation and intelligent

feedback mechanisms, the system aims to improve exercise accuracy and user adherence.

The primary objective of this research is to demonstrate that accurate exercise form classification can be

achieved directly on mobile devices using BlazePose landmarks combined with Support Vector Machine

(SVM) based feature classification, without requiring cloud support [13], [19]. Additionally, the study

explores the use of fine-tuned lightweight large language models to deliver context-aware and safe

physiotherapy recommendations, building upon prior work in medical conversational AI [8], [9], [11].

Another key objective is to validate the feasibility of converting SVM models into mobile-friendly formats

carried out to understand the current advancements, identify existing limitations, and highlight the research

gap addressed in this study. Table I. presents a summary of the key works analysed.

4

Wagh, Scott &

Kraeutne r (2024)

Quantifying similarity

between MediaPipe 2D

upper-limb trajectories

Proof-of-concept

confirming MediaPipe

as viable low-cost

[2] both demonstrating that effective posture correction is achievable without clinical infrastructure. Fourth,

the integration of augmented reality and computer vision for condition-specific rehabilitation (Wang et al.

[7]) signals a trend toward personalised, diagnosis-driven digital therapy, though hardware cost and condition

specificity remain barriers to broad adoption.

The literature review surfaces five critical unmet needs that the proposed TeleRehab AI system directly

addresses:

No existing system integrates real-time pose-based exercise monitoring with a fine-tuned large language

model capable of generating structured, clinically grounded prescriptions in natural language. All reviewed

inference, and longitudinal session tracking within a single coherent platform deployable without specialist

hardware [1, 3, 4, 5].

The surveyed studies reveal four dominant trends across recent telerehabilitation research. First, webcam and

RGB-camera systems have largely displaced depth-sensor approaches as the default hardware choice, driven

by the validation of MediaPipe BlazePose as a sufficiently accurate substitute for expensive motion-capture

equipment [3, 4]. Second, exercise monitoring systems consistently prioritise

Physiotherapy plays a vital role in the recovery process for individuals suffering from injuries, post-

fitness trainers. However, such supervision is not always consistently available. Although telerehabilitation

systems have been introduced to overcome these challenges, many of them rely on continuous internet

access, live video monitoring, or specialized wearable devices. These requirements increase system

complexity, cost, and limit their applicability in resource-constrained settings.

In addition, most existing solutions do not offer real-time feedback or adapt their guidance based on

individual patient conditions. The lack of intelligent monitoring and personalized recommendations often

leads to reduced patient engagement and lower adherence to prescribed exercise routines, contributing to

higher dropout rates.

The fundamental design decision to process all pose data on-device was driven by three critical requirements:

The model performs well across all three squt classes, with each showing a distinct trade-off. For Correct

squats, recall is perfect (1.00), meaning no good squats are missed, though precision is lower (0.73) as

some poor-form squats slip through — a clinically acceptable error since false negatives here pose no

injury risk. Forward Bend detection is the opposite: flawless precision (1.00) ensures no good squats are

wrongly rejected, but it misses a third of actual forward-bend cases (recall 0.67), prioritizing user

satisfaction. Shallow squats show the most balanced performance, with both precision and recall at 0.90.

The SVM uses One-vs-One (OvO) multiclass classification with three binary classifiers combined via Platt

scaling [18]. Platt scaling applies a logistic function to SVM decision values to estimate class probabilities

— an operation with no direct TFLite neural-network equivalent. Conversion therefore employed model

TFLite export: Standard TFLiteConverter with DEFAULT optimisations (no post-training quantisation,

full float32 precision retained).

The complete on-device pipeline — ML Kit landmark extraction, knee-angle smoothing, depth window

update,

and TFLite inference — processes frames within the 33 ms budget at 30 fps on mid-range Android

devices. TFLite inference contributes less than 1 ms. Post-rep form verdicts are displayed for 90 frames

( 3 s) to provide sufficient reading time before the next repetition.

Platt scaling can be converted to TFLite using the synthetic grid labeling approach, enabling classical ML

models trained on small physiotherapy datasets to be deployed on mobile devices without rewriting in a neural

framework.

This paper presented TeleRehab AI, a comprehensive mobile physiotherapy assistant integrating a fine-tuned

3Bparameter LLM with on-device SVM-based squat form classification. The system achieves 84.8% leave-

one-out crossvalidation accuracy on a small dataset (33 per-rep examples, 3 classes) and demonstrates that

TFLite models can faithfully replicate SVM decision boundaries via MLP distillation (99.7% agreement, 4.9

KB model size).

The key technical contributions—the latched standing confirmation gate, per-rep worst-case feature

tracking, bilateral landmark occlusion avoidance, and silent background TFLite inference with post-rep

A larger and more diverse dataset will be collected by including multiple subjects and a wider range of

exercises, with a target of at least 200 participants covering all major exercise types. In addition, a

randomized controlled trial (RCT) will be conducted to clinically validate the effectiveness of the system

by comparing outcomes between TeleRehab-guided users, therapist-supervised patients, and unguided

individuals.

Further improvements will include migrating the backend to cloud platforms such as Render or Railway

to enhance scalability and deployment efficiency. Advanced techniques like Dynamic Time Warping

(DTW) will be incorporated to enable continuous form evaluation by comparing user movements with

reference motion sequences. Moreover, the system will introduce pain tracking across sessions to monitor

changes over time, allowing better assessment and communication of patient recovery progress.

3. W. Simoes, L. Reis, C. Araujo, and J. Maia Jr., “Accuracy Assessment ˜ of 2D Pose Estimation with

4. V. Wagh, M. W. Scott, and S. N. Kraeutner, “Quantifying Similarities Between MediaPipe and a Known