Maintenance Notice

Due to necessary scheduled maintenance, the JMIR Publications website will be unavailable from Wednesday, July 01, 2020 at 8:00 PM to 10:00 PM EST. We apologize in advance for any inconvenience this may cause you.

Who will be affected?

Currently submitted to: JMIR Formative Research

Date Submitted: Nov 13, 2026

Warning: This is an author submission that is not peer-reviewed or edited. Preprints - unless they show as "accepted" - should not be relied on to guide clinical practice or health-related behavior and should not be reported in news media as established information.

No-Code Analysis of 52,181 Prison Telehealth Records With a Large Language Model: Formative Case Study of a Natural Language–Only Workflow and Its Failure Modes in Brazil

  • Luiz Fernando Cardoso dos Santos Durão; 
  • Airha de Souza Silva; 
  • Bruna Natália Freire Ribeiro; 
  • Michelle Louvaes Garcia; 
  • Vanessa Queiroz Pereira Cubos; 
  • Roberta Rubia de Lima; 
  • Maria Cristina Coimbra Lages Balestrin de Andrade; 
  • Alexandre Lazinho Santos; 
  • Simone Pacheco Gomide da Silva; 
  • Giovanni Guido Cerri; 
  • Carlos Roberto Ribeiro de Carvalho; 
  • Helder Takashi Imoto Nakaya

ABSTRACT

Background:

Large administrative health databases are often underused because the people who need answers depend on technical staff to translate operational questions into code. Dashboards expose predefined indicators but cannot answer new questions. Large language models (LLMs) can generate executable code from natural language and may offer a complementary interface, provided their outputs are checked and remain auditable.

Objective:

To test whether a natural language–only ("no-code") workflow, in which the operator never writes or edits code, can convert operational public health questions into reproducible analytical outputs from a real-world prison telehealth dataset; to document the errors that arise and how they were detected; and to make the process publicly inspectable.

Methods:

We applied a no-code workflow to 52,181 anonymized consultation records (293 variables) from TeleSAP, the prison telehealth program of São Paulo State, Brazil (June 2024–July 2025). One operator interacted with Gemini (Google) through its web interface in March 2026, in a single thread of 360 prompts, to obtain Python scripts for 17 management questions, executed in Google Colab. Execution errors and implausible numbers were returned to the model as corrective prompts; code was never edited by hand. Gemini then generated a public portal and reconstructed, from each final script, the prompt that reproduces it.

Results:

Of 17 scripts, 15 ran and answered their question; 1 was excluded because the model silently substituted simulated values for the absenteeism field; 6 were curated into public case studies (chronic disease burden by age, teleconsultation duration, first-contact resolution, gender identity and clinical complexity, respiratory seasonality, cardiovascular risk). Illustrative outputs: hypertension in 39.9% of consultations at age ≥45 years vs 11.3% under 45 (P<.001); 64.8% of consultations resolved at first contact; respiratory peak in August (7.9%; P=.008). Of the 360 prompts, 47 returned an execution error to the model, about 20 challenged a number the operator judged implausible, and about 80 adjusted figures or layout. Five failure modes were documented, none detectable from charts: simulated data presented as real; hard-coded climatological temperatures presented as observed; keyword mentions counted as consultations; 5 different denominators across 6 scripts; and results that changed with the operational definition (first-contact resolution 90.2% vs 64.8%; respiratory peak December, P=.40, vs August, P=.008; high cardiovascular risk 40.7% of 592 measured vs 8.6% of 22,286 patients). The portal (telesap.inovalab.tools) presents each case study with prompt, code, and output.

Conclusions:

A natural language–only workflow can turn management questions into working, inspectable analyses of a real dataset, but the dominant risk is a plausible wrong answer: the model fabricates inputs, embeds unstated assumptions, and implements whichever definition it is given. Human verification of inputs, denominators, and definitions, with versioned prompts, code, and model identifiers, is a precondition for LLM-assisted public health analytics.


 Citation

Please cite as:

Durão LFCdS, de Souza Silva A, Freire Ribeiro BN, Louvaes Garcia M, Queiroz Pereira Cubos V, Rubia de Lima R, Coimbra Lages Balestrin de Andrade MC, Lazinho Santos A, Pacheco Gomide da Silva S, Guido Cerri G, Ribeiro de Carvalho CR, Imoto Nakaya HT

No-Code Analysis of 52,181 Prison Telehealth Records With a Large Language Model: Formative Case Study of a Natural Language–Only Workflow and Its Failure Modes in Brazil

JMIR Preprints. 13/11/2026:116900

DOI: 10.2196/preprints.116900

URL: https://preprints.jmir.org/preprint/116900

Download PDF


Request queued. Please wait while the file is being generated. It may take some time.

© The authors. All rights reserved. This is a privileged document currently under peer-review/community review (or an accepted/rejected manuscript). Authors have provided JMIR Publications with an exclusive license to publish this preprint on it's website for review and ahead-of-print citation purposes only. While the final peer-reviewed paper may be licensed under a cc-by license on publication, at this stage authors and publisher expressively prohibit redistribution of this draft paper other than for review purposes.