Authorized input
Simulated, synthetic or properly de-identified audio is introduced into the research pipeline.
ISDAM Clinical AI · Research & Development
Researching safer AI for psychiatric documentation.
ISDAM Clinical AI is a Brazilian research initiative exploring how authorized clinical conversations can become accurate transcripts and structured documentation drafts in Brazilian Portuguese—with mandatory professional review.
Brazilian Portuguese
Human review required
Synthetic or de-identified data
Pre-clinical R&D
The research problem
Clinical conversations are rich. Documentation is demanding.
Psychiatric encounters depend on nuance, context and careful language. Converting these conversations into complete records can require substantial time, while general-purpose transcription systems may miss specialized terminology, speaker changes and the structure expected in mental status examinations and progress notes.
The research question is not whether AI can replace clinical judgment. It is whether AI can produce safer, traceable and useful drafts for qualified professionals to review.
Proposed research workflow
From authorized audio to a reviewable draft.
The prototype separates speech processing, structured generation and professional review so that each stage can be evaluated independently.
Speech processing
Speech-recognition models are evaluated for Brazilian Portuguese, psychiatric terminology, timestamps and speaker separation where available.
Structured drafting
Language models organize the transcript into draft Mental Status Examination, SOAP and clinical progress-note formats.
Indicator extraction
The prototype may identify documented elements related to suicidal ideation or agitation for structured professional review. It does not calculate definitive risk or recommend treatment.
Human review
Every output remains an editable draft. A qualified professional must verify the transcript, correct errors and approve any resulting documentation.
Evaluation framework
Measured as a research system, not marketed as a finished product.
Transcription accuracy
Word error rate, terminology errors, speaker attribution and robustness to realistic background noise.
Clinical fidelity
Omissions, unsupported statements, contradictions and preservation of clinically relevant context.
Structured extraction
Precision, recall and consistency for predefined documentation fields and research indicators.
Operational feasibility
Inference latency, token and audio costs, reproducibility and performance across model configurations.
Human review
Correction burden, edit distance and professional assessment of whether drafts are complete and reviewable.
Planned cloud evaluation
Evaluating speech and language models on Alibaba Cloud.
ISDAM plans to evaluate Alibaba Cloud Model Studio, including Qwen speech-recognition and text-generation capabilities available for international workloads. The proposed study will compare inference configurations using synthetic or properly de-identified material and measure transcription quality, factual consistency, safety, latency and cost.
This describes an intended technical evaluation only. ISDAM is not currently sponsored, endorsed or officially partnered with Alibaba Cloud.
Speech-to-text inferenceStructured note generationPrompt and model comparisonSafety and hallucination testingLatency and cost measurement
Responsible AI
Clinical sensitivity requires explicit boundaries.
Pre-clinical status
The project is a research prototype and is not available for clinical use.
Data minimization
Initial development uses synthetic, simulated or properly de-identified material. Identifiable patient information must not be submitted.
Human oversight
AI outputs are drafts. They require verification, correction and approval by a qualified professional.
No autonomous decisions
The system must not diagnose, prescribe, determine disposition or independently classify psychiatric risk.
Ethical gates
Any future research involving participants or identifiable clinical information depends on applicable ethical, institutional, legal and data-protection review.
Transparent limitations
Model errors, omissions, bias and hallucinations are treated as central evaluation outcomes rather than hidden limitations.
Research roadmap
A staged path from benchmark to evidence.
Current focus
Synthetic benchmark design
Define simulated psychiatric dialogues, output schemas, evaluation criteria and safety tests.
Model comparison
Compare speech-recognition and language-model configurations for quality, latency, consistency and cost.
Prototype hardening
Improve traceability, output validation, access controls, logging and mandatory review workflows.
Ethical evaluation
Any study involving human participants or clinical data will begin only after the required approvals and safeguards are in place.
About ISDAM
Medical experience guiding responsible clinical AI research.
ISDAM is the registered trade name of Instituto de Saude Doutor Andre Moura Ltda, a Brazilian medical-services microenterprise founded in 2023 and headquartered in Araguaína, Tocantins. ISDAM Clinical AI is its research and development initiative focused on responsible applications of artificial intelligence in clinical documentation.
The research direction is led by Dr. André Junior Francelino de Moura, CRM 13788-MA, a physician and third-year Psychiatry Resident at Hospital Nina Rodrigues in São Luís, Maranhão. The project is connected to his medical-residency thesis work on artificial intelligence and psychiatric documentation.
Professional affiliation is provided solely as biographical context and does not imply institutional sponsorship, partnership, approval, validation or endorsement of this project.
- Legal name
- INSTITUTO DE SAUDE DOUTOR ANDRE MOURA LTDA
- CNPJ
- 51.403.920/0001-91
- Status
- Active
- Location
- Araguaína, Tocantins, Brazil
Contact
Research and technical collaboration.
For research, technical or institutional inquiries, contact ISDAM through its corporate email.
diretoria@isdam.com.brDo not send patient information, medical records, recordings or other clinical data by email.
Technical and governance references