A system developed at the West University of Timișoara shows how medical screening can become faster, more accurate, and fully secure in terms of data privacy.
Why early autism diagnosis is essential
Early diagnosis of autism spectrum disorders is not merely a formal medical step, but a critical moment in a child’s development. The earlier therapeutic intervention begins, the greater the chances of adaptation, communication, and social integration.
In this context, technology is becoming an important ally of modern medicine. Artificial intelligence (AI) is beginning to be used not only for automation or data analysis, but also to support clinical decision-making.
A new research project developed at the West University of Timișoara shows how AI can contribute to autism screening in a fast, efficient, and, very importantly, fully secure way from the perspective of patient data protection.
A medical innovation developed in Romania
The research is carried out by Alexandru Robert Vlasiu and Marc Eduard Frincu, as part of a project funded through the grant “Counselling and Career Guidance Centre for Researchers – Western Region”, supported by NRRP funds from the Ministry of Research.
Their objective is clear: to develop an intelligent screening system for the early detection of autism spectrum disorders, combining the analytical accuracy of artificial intelligence with high standards of medical data protection.
The proposed solution is not just a classification algorithm, but a complex system designed to function in real clinical environments, where data confidentiality and security are essential.
How the system works: two levels of intelligence
The architecture proposed by the West University of Timișoara researchers is hybrid, built on two main components that work together.
1. Data protection through the PATE method
The first component is based on an advanced data protection method called PATE — Private Aggregation of Teacher Ensembles.
In simple terms, the system works like a group of “digital experts” who analyse patient data separately. Each “teacher” learns from a subset of data without having access to the complete identity of the patients.
The final decision is not made by a single model, but by aggregating the results of all the teachers. In this process:
- personal data remain fully anonymised;
- sensitive information does not leave the source institution;
- the final result is built statistically, not individually.
This approach allows AI to be used in medicine without compromising patient confidentiality, an essential aspect in the context of European data protection regulations.
2. Clinical analysis through advanced language models
The second component of the system is an analysis module based on natural language artificial intelligence models, capable of interpreting clinical descriptions.
This module, conceptually called MedPrompt, does not simply look for keywords, but analyses the behavioural context described by physicians.
For example, instead of identifying terms such as “social difficulties” or “repetitive behaviour” in isolation, the system analyses the relationships between these observations and builds a global assessment of the case.
In practice, it functions similarly to a panel of virtual specialists who discuss and interpret a clinical case together.
High performance in standardised screening
During the testing phase based on the AQ-10 questionnaire, the system demonstrated very solid results.
The model achieved:
- high accuracy;
- maximum sensitivity and specificity in the evaluated dataset;
- stable results even under data protection conditions, through the introduction of statistical noise.
These results suggest that the private aggregation mechanism works effectively without affecting the quality of the final decision.
However, the researchers emphasise that these performances were obtained in a controlled environment, where the data structure is well aligned with the AQ-10 instrument.
AI analysis on narrative clinical profiles
An important component of the study was the evaluation of the system on narrative descriptions of clinical cases.
In this scenario, the AI model achieved:
- overall accuracy: 94%;
- sensitivity: 88% — correctly identifying autism cases;
- specificity: 100% — correctly excluding non-autism cases.
This combination indicates a highly precise system, with a strong ability to avoid false positives, which is essential in a medical context.
The observed errors occurred especially in “borderline” cases, where symptoms are more subtle and may be confused with normal variations in development.
The importance of context in AI performance
A key result of the research was related to scenarios in which the system had incomplete information.
When AI was used based only on behavioural observations:
- accuracy was 65.4%;
- specificity was low, at 46.2%.
However, when the researchers added a single contextual variable — the child’s age — performance increased dramatically:
- accuracy: 88.5%;
- specificity: 92.3%.
This difference highlights an essential point: medical artificial intelligence needs minimal clinical context in order to avoid erroneous interpretations.
In other words, AI does not function optimally in isolation, but in combination with basic medical data.
Robustness tests and validation on synthetic data
To verify the stability of the system, the researchers carried out stress tests using synthetic data, with complex variations and rare combinations of symptoms.
Under these conditions:
- the system’s performance decreased slightly, from 100% to approximately 97.2%.
This decrease is considered normal and indicates that the model is not “overtrained”, but capable of generalising well under varied conditions.
However, the researchers underline that the system remains a research prototype, which requires extensive clinical validation before any practical use in medical diagnosis.
Artificial intelligence in medicine: between promise and responsibility
The project developed at the West University of Timișoara highlights an important aspect of the future of medicine: the use of AI does not replace the physician, but supports them.
Systems of this type can:
- accelerate the screening process;
- reduce pressure on the medical system;
- improve the accuracy of initial assessments;
- protect patients’ sensitive data.
At the same time, they must be used responsibly, in strictly controlled and scientifically validated contexts.
Impact for Romania and the medicine of the future
The development of this type of technology in Romania, within a project funded through the NRRP, shows the potential of local research to contribute to global innovation.
For the medical system, such solutions can open the way towards:
- faster screening for developmental disorders;
- reduced diagnostic time;
- better access to specialised assessments in areas with limited resources;
- responsible use of artificial intelligence in healthcare.
For the scientific community, the project demonstrates that AI can be integrated into sensitive fields such as paediatric medicine without compromising ethics or data security.
Conclusion: a step towards smarter and safer medicine
The research carried out at the West University of Timișoara shows that the future of medicine is not only digital, but also responsible.
By combining artificial intelligence with advanced data protection methods, the researchers have demonstrated that it is possible to build systems capable of supporting the early diagnosis of autism without compromising patient confidentiality.
It is a clear example that technological innovation, when guided by ethical principles and scientific rigour, can have a real impact on people’s lives.
Scientific article
The research can be consulted here:
https://doi.org/10.1016/j.actpsy.2026.106996
“The content of this material does not necessarily represent the official position of the European Union or of the Romanian Government.”
“NRRP. Funded by the European Union – NextGenerationEU”
Website: https://mfe.gov.ro/pnrr/
Facebook: https://www.facebook.com/PNRROficial/