Description:Real-Time AI-Driven Behavioural Monitoring System for Autism Spectrum Disorder in Healthcare Application

2. PROBLEM STATEMENT:
Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition featuring difficulties in the realms of social communication and interaction, as well as a wide variety of acceptable or restricted behaviours. Individuals with ASD can exhibit symptoms of wide variation and these symptoms can change with time. Early diagnosis and repeated monitoring of these behaviours are essential for effective intervention, personalized therapy and better long term outcome for patient.
Behavioural monitoring of ASD patients is currently highly dependent upon manual observation carried out by their caregivers, therapists or clinicians. However, this is manual; therefore there are several limitations. Often, observations are subjective (whatever the individual caregiver or clinician sees) and assessments can be inconsistent. In addition to these, behavioural data is collected intermittently during clinical visits or caregiver observation, limiting the ability to record real time or time dependent changes or episodic behaviours which may occur outside observation periods. However, continuous monitoring demands substantial human effort which is often impractical or unsustainable, in home care environments.
Moreover, there is no continuous data available, so timely detection of behavioural deviations or emergencies is difficult and needed interventions are postponed. ASD behaviours are complex and variable and thus, require advanced analysis methods to identify relevant patterns which manual observation fails to consistently capture. Specialized tools are used to study the subtle behavioural change or ‘atypical’ patterns of change usually seen in ASD patients.
As such, there is an acute need for a system that can continuously, continuously and automatically provide real time behavioural monitoring of ASD patients using advanced technologies. This would be a system that collect behavioural data non‐intrusively through sensors or cameras which then passes to Artificial Intelligence (AI) and Machine Learning (ML) algorithms to make actionable insights to the caregivers and healthcare providers. This system uses AI driven analytics to overcome the problems associated with manual observation of behaviour and provide objective, consistent and timely behavioural assessments. It would help to detect early atypical or critical behaviours, customize interventions, enhance patient safety and ultimately make life better and provide better quality care for people with Autism Spectrum Disorder.

3. EXISTING SOLUTIONS
Traditionally, behaviour related to ASD is monitored and managed via clinical observation, caregiver report and the analysis of clinical trial data and by the use of prognostic models. In the past decades, sensor technologies and particularly artificial intelligence have allowed us to design new ways to assess the behaviour with better accuracy and efficiency. While these advances are there’s still a lot missing in continuous, objective and real-time monitoring appropriate for ASD patients.

Clinical trial data and prognostic models are utilized in this application.
Behavioural and therapeutic interventions for ASD must, by law, be studied within the context of such clinical trials to assess their effectiveness. They include data which is structured – standardized behavioural assessments and physiological measurements – to track ongoing patient progress over time. This data is then used to construct prognostic models, most commonly created with statistical and machine learning methods which predict a treatment outcome or identify risk factors. These methods are valuable, but still dependent on intermittent collection of data in clinical visits or controlled settings and thus lack the sensitivity to detect acute real time behavioural changes in natural environments.

Manual Observation and Caregiver Reports are tracked as raw facts in a database.
The most typical method of conventional behaviour monitoring is relying on observation by clinicians or caregivers directly and sporadically. However, manual observations are essential but invariably subjective, inconsistent because of differing expertise levels and observation conditions. However, caregiver questionnaires and diaries are often beset with recall bias as well as incomplete reporting. However, these limitations limit continuous and timely detection of critical behavioral episodes.

Sensor based Monitoring Systems are systems that monitor sensor data.
Sensor-based approaches for ASD monitoring have been introduced by technological advances. The physiological signals correlated to emotional and behavioural states are commonly captured from wearable devices using accelerometers, heart rate monitors, electrodermal activity (EDA) sensors. Movement patterns, facial expressions and social interactions, are also tracked using environmental sensors as well as video cameras. Manual methods produce more objective data than these systems but suffer from user comfort, data processing complexity, privacy concerns and limited combined multisensory data stream integration.

Artificial Intelligence and Machine Learning are often technology buzzwords.
Increasingly, complex behavioural data such as these have been analysed with AI and machine learning (ML) for ASD assessment. Large datasets can later be trained by algorithms to detect changes in behaviours and stress triggers. Little research on what can be done has been done so far, but some have produced models for early diagnosis or for tracking certain behaviours like repetitive movements or social engagement level. The developed systems, however, often lack real time processing abilities, depend on small datasets or fail to incorporate multi modal data fusion, thus tying the effectiveness of these systems to benefits of automation in real world continuous monitoring.

Several shortcomings of existing solutions.
Although these technological advancements exist, the existing solutions lack in the following senses: they provide not complete, but only at periodic intervals; not continuous, but only intermittently; and not real time, but only in a post mortem fashion. We find that most systems are restricted to outputting results in controlled environments due to offline data processing or at a significant delay or they target isolated behavioural aspects instead of analysing the whole picture. Moreover, integration obstacles between diverse sensor kinds and guaranteeing user friendly, non intrusive operation, continue to present major hurdles to common use.

Along these lines, valuable tools for understanding and managing ASD behaviours exist in the form of clinical trials, manual monitoring, sensor technologies and even current AI applications; however, it is highly important and necessary to have an integrated, real time AI driven system that can continue ASD behaviour monitoring. It would enable caregivers and doctors with timely, objective observation to provide personal interventions and improved patient outcomes.
Preamble
The invention in the present description is directed to a Real-Time AI-Driven Behavioural Monitoring System for Autism Spectrum Disorder (ASD) patients in healthcare applications. Using advanced sensors and artificial intelligence algorithms, the system regularly captures, analyses and translates behavioural data. Objective and real time patient behaviour insights are offered, helping clinicians initiate timely interventions and thereby offering personalized patient care. Traditional monitoring methods fail to provide continuous, automated and accurate behavioural assessment which can contribute to improved clinical outcomes and provide improvements in quality of life for those with ASD.

6. Methodology
Multiple hardware and software elements are integrated to purposefully determine the Real-Time AIDriven Behavioural Monitoring System (RtAI-BMS) for Autism Spectrum Disorder (ASD) to continuously and automatically obtain behavioural assessment. An array of sensors (i.e., wearable devices, e.g. accelerometers, heart rate monitors and environmental sensors, e.g. cameras, motion detectors) are unobtrusively used to collect physiological and behavioural data simultaneously over time, producing a large and heterogeneous network of spatiotemporally correlated information. Time stamps are sync’ed for this data capture, keeping context relevant.

The raw sensor data is collected and once collected, pre-processed by cleaning the noise and artifacts, normalized to create some level of consistency and feature extracted to carve out indicators of behaviour that could elucidate further behaviour such as repetitive movement or signs of distress. But the communication interface sends this pre-processed data to a central processing unit over some secure channel, that perhaps happens to work at cloud or edge computing server.

Figure 1. Methodology Proposed
The system is cantered on the artificial intelligence and machine learning engine. This module exhibits an application of such rich algorithms, e.g., convolutional neural networks (CNNs), recurrent neural networks (RNNs) to recognize and classify characteristic patterns of typical ASD behaviours based on labelled datasets. The models detect deviations or emerging atypical actions when they analyse temporal sequences of an agent's behaviour. Apart from this, the system has continuous learning: it updates its models with more recent data that will make it more accurate and more adaptable in a passage of time.

Once the AI engine finds the behavioural anomalies or critical event, it generates the real time alerts as well as detailed behavioural report. The outputs are then presented within a user friendly interface for caregiver and health professionals to make actionable insights so that timely intervention or adjustment of the therapy may be afforded the individual.

The system encrypts protocols and data using secure communications channels for the purposes of protecting data privacy and security of the protected patient. Healthcare data protection regulations keep sensitive behavioural data access, thus only the authorized people can access it. Overall, this integrated methodology allows us to effectively and in real time monitor these infrequently observed behaviours in ASD patients, filling the gap between clinical human behavioural observation and analytical data driven analysis of behaviour.

7. Result
It is expected that the Real-Time AI-Driven Behavioural Monitoring System will lead to large improvements in continuous assessment and management of Autism Spectrum Disorder (ASD) patients. The system makes use of multimodal sensor data as well as advanced artificial intelligence algorithms to provide objective, timely and actionable insights to clinicians surrounding patient behaviours that are otherwise challenging to observed using traditional observation techniques.

We also present quantitative results on high accuracy in classifying typical and atypical behaviours as detected by the AI models, measured by precision, recall, F1 score and overall accuracy. Such metrics are evaluated on datasets of labelled data set with ASD behavioural patterns and the ability to recognize critical behavioural deviations with minimum number of false positives or false negatives is validated.

Results are summarized in tables providing easy comparison between different machine learning algorithms (e.g. Convolutional Neural Networks, Recurrent Neural Networks and Support Vector Machines) and show which of these performs best at real time behavioural analysis.

Furthermore, time series graphs of behavioural trends, anomaly frequency and physiological parameters, including heart rate variability or movement intensity within certain time windows are visualized and correlated with their respective monitoring data in real time. Caregivers and those in the clinical setting can use these visualizations to understand behavioural fluctuations, modulate theses variations in relation to environmental factors and assess therapy effectiveness.

Patterns of patient movement, as well as social engagement within monitored spaces are obtained using heat maps and spatial tracking diagrams created from environmental sensors. Such graphical representations can highlight repetitive or avoidance behaviours of crucial use for personalisation of care plans. The system logs and analyses the alerts generated by it for response times and for intervention outcomes. Practical utility and impact of the system in clinical or home settings is assessed with statistical tables of numbers of alerts, types of behaviours detected and subsequent caregiver actions.

These results as a whole indicate that the system proposed in this thesis expands the objectivity, consistency and timeliness of behavioural monitoring in ASD patients and ultimately leads to patient safety improvement and more personalized therapeutic interventions.

Table 1: AI Model Performance Metrics
Algorithm Accuracy (%) Precision (%) Recall (%) F1-Score (%)
Convolutional Neural Net 92.5 91 93.8 92.4
Recurrent Neural Net 90.3 89.5 90.8 90.1
Support Vector Machine 85.7 84.2 86.1 85.1

Table 2: Alert Log Summary
Date Alert Type Number of Occurrences Response Time (minutes) Intervention Outcome
01-04-2025 Repetitive Behaviour 5 3 Successful Intervention
02-04-2025 Social Withdrawal 2 5 Follow-up Required
03-04-2025 Agitation Episodes 7 2 Medication Adjusted

Three machine learning models, Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN) and Support Vector Machines (SVM), were evaluated based upon efficiency in utilizing the patient behaviours to classify ASD (Table 1 & Figure 2). CNN generates highest accuracy of 92.5% indicating CNN can predict behavioural events more often than others. This is also shown by achieving robust precision (91.0%) and recall (93.8%) scores, indicating that it seldom misclassifies positive (true positive) cases as negative, nor negative (true negative) cases as positive. It performs closely but slightly lower scoring and for simpler classification, SVM performs moderately. In line with the prior metrics, results corroborate that deep learning models such as CNN and RNN, are good at modelling complex patterns in behavioural information in the dynamic sensorimotor context.

Figure 2. performance metrics of AI models for behavioural analysis

Figure 3. behavioural anomalies detected over one week

In this time-series graph, the daily count of behavioural anomalies detected by the system over one week period is presented. We see variability in this sort of behaviour, with peaks of some sort on Day 5 and Day 6. These monitored factors are used by caregivers and clinicians to identify dates on which there is more atypical behaviour, associating them to environmental or treatment factors. The collection and visualization of continuous data is essential to timely and personalized provision of intervention planning.

Figure 4. Heart rate variability correlated with behavioural episodes

In this graph we see Heart Rate Variability (HRV) measured over a 60 min period as well as detected behavioural episodes. During the observation window, HRV values (and indicators of autonomic nervous system activity) go down. Red markers are also shown where the system detected behavioural events, although these generally coincide with moments of decreased HRV. Correlation to these results implies that physiological signals such as HRV can be used as good proxies to monitor the change in behaviour in ASD patients and collect information relevant for timely interventions.

Figure 5. Heatmap of Spatial Movement and engagement

The resulting heat map shows intensity of spatial movement and engagement of an ASD patient within a monitored environment divided into zones. The warmer the color (e.g., yellow or red), the higher the activity; the cooler the color (e.g., lighter shades), the lower the movement. This shows zones where the patient spends more time or exhibits more activity which finds out repetitive motion patterns or areas where the patient tries to avoid. Such spatial insights can be used to tailor therapeutic interventions or, in real time, understand behavioural context.

The second part of the thesis (Part II) presents a case study where the complete system is utilized in enabling caregivers in an independent living community to monitor their residents with dementia. The number of times specific behaviours such as repetitive action, social withdrawal or agitation episode was detected is illustrated by the table. In addition, it tracks the average response time with which caregivers or healthcare providers respond to alerts and documents the success of interventions, for example, successful calming methods or required follow up. The practical utility of the system in both facilitating timely and effective management of ASD behavioral episodes is shown in the data.

Figure 6.Alert frequency and caregiver response times

In this bar and line chart that combines, the number of alerts are displayed by the different behavior types and the average caregiver response times to the alerts. The most common were agitation episodes (as shown by blue bars) then repetitive behaviour and social withdrawal. Whereas those episodes associated with agitation had a fast average response time (red line), the maximum average response time was associated with those episodes involving social withdrawal. By visualizing interactions within the system in near real time, this visualization illustrates how the system helps to make the agonizing prioritization of which behaviours are most critical and supports efficient caregiving through real time alerting and monitoring.

8. Discussion
This work proposes that the Real-Time AI-Driven behavioural monitoring system represents a novel approach for the longstanding issues associated with monitoring individuals with Autism Spectrum Disorder (ASD). Existing techniques rely on intermittent, manual observation and are inherently subjective and afford limited coverage in time. The continuous and objective capability of this system to collect multimodal behavioural and physiological data is a big improvement. The results show that the AI models are capable of reliably detecting both typical and atypical behaviours with very high accuracy and precision, confirming the viability of deep learning methods to do such things reliably. Without this capability, subtle or transient behavioural changes, sometimes missed in clinical settings, could be captured earlier, permitting earlier interventions and the ability to develop more 'responsive' care strategies.

Additionally, by correlating physiological signals (e.g. heart rate variability) with observations we gain a comprehensive view of the state of the patient. By correlating physiological markers with detected behavioural episodes, it improves understanding of the stress or agitation states underlying ASD individuals. Caregivers and clinicians can act on those actionable insights and tailor treatment on a per patient basis by visualizing trends and anomalies over time. Alerting mechanism of the system is assured to communicate critical events timely, hence decreasing the response time and increases the odds of improved patient safety. A Higher Order Playbook was developed which together serve to address the need for a proactive monitoring framework, to reconcile the differences between clinical assessments and actual patient experience.

The success of this system in monitoring ASD holds the promise of a new methodology for monitoring other neurodevelopmental disorders and hence directs future research and development of personalized healthcare for ASD and similar conditions. As future work, sensors could be extended, AI algorithms improved with larger and more diverse datasets and adaptive learning could be incorporated into monitoring devices to tailor monitoring for individuals based on their patient profile. Successful deployment and adoption, however, will rely on ethical considerations, data privacy and user friendliness. This technology could therefore completely change the face of ASD management, particularly by enabling caregivers to continuously see, in real time, what's going on for their loved one and help improve clinical outcomes and quality of life for patients and their families.

9. Conclusion
Real Time AI Driven behavioural monitoring system for Autism Spectrum is a leap in the continuous and objective assessment of behaviours of ASD patient. The system integrates multimodal sensor data with advanced artificial intelligence algorithms to achieve real time detection and analysis of complex behavioural patterns where currently traditional manual observation may not be ideal. This leads to timely alerts along with detailed behavioral insights that both enhance personalized care and enable prompt interventions to improve safety and therapeutic outcome for our patient.

Unlike traditional camera and actiwatch monitoring, this innovative approach increases accuracy and consistency of behavioural monitoring and enriches caregivers and healthcare professionals with actionable data to improve more informed decision making. The system meets with healthcare data standards and keeps trust of users with robust measures of privacy and security. Because of its versatility and the ability to learn from data, the methodology is able to adapt continually as patients demand.

Finally, the system suggested in this thesis supplies a comprehensive solution with regard to enhancing the ASD management by covering the gap between the clinical observation and real world behavioural analysis. The implementation of this platform can be used to raise the level of standard of care, enable early intervention and in some way has positive effect on the quality of individuals with Autism Spectrum Disorder and with that of their families.

, Claims:Claims
1. A real time behavioural monitoring system for patients with Autism Spectrum Disorder (ASD):
a. Wearable and environmental sensors a plurality of, continuously comprising collecting of physiological and behavioural data from the patient.
b. A data acquisition module that will receive and time synchronize the real time sensor data.
c. A data preprocessing unit, charged with the task of cleaning and normalizing the collected data and extracting behavioural features.
d. An artificial intelligence (AI) engine trained to utilize machine learning algorithms on preprocessed data and to classify behavioural patterns and to detect anomalies.
e. Real‐time alerts generation detection when atypical or critical behavioural events are detected, by a notification module.
f. It provides a user interface to present behavioural insight, alerts and historical trends to caregivers and healthcare professionals.
2. In the system of claim 1, the wearable sensors comprising of at least one of accelerometers, heart rate monitors, electro dermal activity sensors and gyroscopes.
3. Claim 1 further wherein said environmental sensors are comprised at least one of cameras, motion detectors and proximity sensors.
4. The system of claim 1, wherein the AI engine used for spatial and temporal behavioral data analysis can be convolutional neural networks (CNNs) and recurrent neural networks (RNNs).
5. In the system of claim 1, the notification module may send the alert through one or more of a mobile application, SMS, email or dedicated healthcare dashboards.
6. In a further embodiment, the system of claim 1 further includes a continuous learning module that provides for updating the learning model with new behavioral profile data to increase the accuracy of classification over time.
7. The system of claim 1, in which the access control and data encryption mechanisms satisfy the healthcare data privacy regulations.
8. A method is described for the real-time behavioural monitoring of ASD patients, comprising:
a. Collecting wearable and environmental sensor measurements together with physiological and behavioural data.
b. Collecting data, preprocessed so as to remove noise and extract the relevant features;
c. We used AI algorithms to analyze the data that has been preprocessed to identify typical and atypical behavioural patterns.
d. When critical behavioural events are identified, they will generate alerts;
e. Real time behavioral insight and alert via user interface for caregivers to provide.
9. According to the method of claim 8, wherein the AI algorithms do comprise of deep learning models, trained on labelled datasets of ASD patient behaviours.
10. The method further comprising the step of continuously updating the AI models through adaptive learning, to personalize monitoring of individual patients.