Log inSchedule a demo
Menu

Research

Science and evidence

Software for the mental health record must be demonstrably careful. On this page we keep three things apart: what we know about Notal Care itself, what research on language models in general does and does not show, and which research we are preparing.

01Current position

Where we stand.

Notal Care is academically validated and scientifically anchored at LIACS (Leiden University), where our research method was also validated. The product has been validated in practice at In-Dividu (an independent mental health organisation) and Praktijk Rood (youth and adult care).

  • Transcription, summaries and case formulation

    In use in practice

  • Querying the record

    First version, has limitations

  • Research into the quality of draft reports

    Running since October 2026

  • Academically and practice-validated

    Completed, follow-up running

  • Clinical modules in MDR certification, class IIa

    Expected Q3 2027

Status as of October 2026

02Collaboration

Collaboration with Leiden University.

The collaboration with Leiden University takes shape through Dr Niki van Stein, associate professor at the Leiden Institute of Advanced Computer Science (LIACS), where she researches automated machine learning, global (Bayesian) optimisation and neural architecture search. LIACS has validated our research method. Supervision of thesis research has been agreed; interdisciplinary research with the psychology department is planned. For each study we set out in advance who takes part, which question, which method and which data. We publish results, including disappointing ones.

03Research on Notal Care

What we are researching about Notal Care itself.

For each study: the question, the design, the setting, who carries it out and what it does and does not show.

Ongoing research, since October 2026

Quality of draft reports for Dutch mental health sessions

Question
How good are Notal Care's draft reports for Dutch mental health sessions?
Design
Research method validated by LIACS. Design, assessors and outcome measures are set out in advance.
Setting
Dutch mental health sessions
Conducted by
Notal Care, supervised by LIACS (Leiden University)
What it does and does not show
No results yet. Until then we make no claims about quality, time savings or correction effort.

Validated in practice

Validation in practice with the first customers

Question
How do transcription, summaries and case formulation work in clinicians' daily work?
Design
Validation in practice at the first organisations; not a controlled study.
Setting
In-Dividu (independent mental health organisation); Praktijk Rood as an additional practice setting for youth and adult care
Conducted by
Notal Care with the first customers
What it does and does not show
Practical experience in a defined setting. Says nothing yet about how it works with other target groups, forms of treatment or organisations.

In certification, expected Q3 2027

Clinical evaluation of the clinical modules

Question
Are the clinical modules safe and do they do what they promise?
Design
Clinical evaluation as part of MDR certification (class IIa).
Conducted by
Notal Care, as part of MDR certification
What it does and does not show
Until CE marking, these modules are not available for clinical use.

04External literature

What research on language models does and does not show.

These studies examine language models and speech-driven documentation in general. They inform choices in our development, but none of these studies tests Notal Care itself.

External study

Attention during the consultation

Question
What do psychiatrists experience with speech-driven documentation?
Design
Experience study in simulated consultations
Setting
Eight psychiatrists and four simulated patients
Conducted by
Bokhari et al. (2026)
What it does and does not show
Participants described less distraction from documentation and more room for contact, alongside concerns about confidentiality, dependence and loss of their own clinical wording. Simulations; no demonstrated improvement of care in practice.

External study

The transcript is also an edit

Question
Does speech recognition sometimes add text that was never spoken?
Design
Analysis of speech recognition ("Careless Whisper")
Conducted by
Koenecke et al. (2024)
What it does and does not show
The speech recognition studied sometimes added sentences that were not spoken, more often during long silences and with aphasia. This is not an error rate for current models or for Notal Care; it underlines that the transcript must be checked separately.

External study

Model performance and the physician's judgement

Question
Does access to a language model improve physicians' diagnostic reasoning?
Design
Randomised study with written cases
Setting
Fifty physicians
Conducted by
Goh et al. (2024)
What it does and does not show
No statistically significant improvement in the physicians' reasoning, while the model performed strongly on its own. What the technology can do and what it adds to the professional's practice are different questions.

External study

Judge first, then compare

Question
How does structured collaboration between clinician and AI work out in diagnostic tasks?
Design
Independent assessment by human and AI, followed by comparison and discussion
Setting
Seventy clinicians; vignettes
Conducted by
Everett et al. (2026)
What it does and does not show
Favourable results on diagnostic case tasks, but also mutual influence and cases in which AI made the judgement worse. Vignettes; no real treatment contacts.

05Planned research

Three questions on which we seek collaboration.

Each question needs its own expertise: technical reliability, clinical quality and insight into meaning. For each question we are looking for researchers and care practices who want to contribute.

  • Reliable processing

    Which clinically relevant errors arise between session, transcript, summary and record, and are corrections retained?

  • Professional and relational quality

    What does use mean for judgement, continuity and perceived attention in the session, with an acceptable checking burden?

  • Meaning and transparency

    What context is lost, and when does the presentation wrongly influence professional judgement?

06People

Advisory board

Vincent Bol is a director of In-Dividu, one of Notal Care's first users, and advises Notal Care on clinical development.
To the advisory board

Vision: position paper

Why sustainable AI in mental health care must be clinically embedded, governance-driven and grounded in practice.
To the position paper

07Questions

Frequently asked questions.

Is Notal Care scientifically validated?

Yes. Notal Care is academically validated and scientifically anchored at LIACS (Leiden University), where our research method was also validated. The product has been validated in practice at the first Dutch mental health care (GGZ) organisations using it. Follow-up research has been running since October 2026; we publish the results on our science page.

Do the studies mentioned prove that Notal Care works?

No. They examine language models and speech-driven documentation in general, each with their own tasks, participants and limitations. They help us choose what to develop and what to test. Whether Notal Care itself works has to be shown by research on Notal Care.

Is Notal Care a medical device?

Partly. The core for documentation and record building is not. The clinical modules are being certified as a class IIa medical device; they will become available after CE marking, expected in the third quarter of 2027.

Does Notal Care make clinical decisions?

No. Notal Care organises session material and makes suggestions; you assess, edit and decide. As development continues, clinical judgement stays with the clinician.

Would you like to take part in research?

For care organisations that want to take part in an evaluation and researchers who want to work on one of the three questions.

Schedule a conversationSchedule a demo
Schedule a demo