The fundamental idea of this paper is to provide an alternative for reducing the time dedicated to the tedious process of writing clinical reports while simultaneously generating added value.

Time is the most valuable asset we have, as it cannot be recovered. Therefore, it is essential to invest time in the best possible way. In modern medicine, with a heavy care burden, we spend a significant part of our time writing clinical reports and filling in databases.

However, we know that clinical reports per se do not harbor much value, and our ‘administrative’ work is filled with inefficiencies by entering the same parameters multiple times (e.g., patient admission dates in their medical history, RETRAUCI,1 and ENVIN2). However, records generate value since we manage to group the information, facilitate statistical analysis, and consequently generate knowledge.

If the previous observation is correct, we should spend more time (being a scarce and finite good) on what brings more value, the registries. However, this would leave us with no time for clinical reports.

Currently, there are natural language models based on artificial intelligence (MLN). These models could help us save time in writing discharge reports, allowing us to spend time creating quality records in registries. Medical knowledge plays a significant role in the quality of the records as it serves as a verifier of the information entered in difficult-to-classify patients.

The fundamental idea is to introduce copious and quality information into the national registries, export a structured report from national registries, and from there ‘predict’ the discharge report through artificial intelligence using MLN.

What is proposed here is an inversion of medical activity. We propose that medical information shift its priorities and gradually prioritize filling out records - since this is where we condense the value - and from these records, we export a structured text on which we can apply natural language models to 'predict' the clinical report.

MLNs work with input information (structured report generated from registries), a determined algorithm, ‘prompt’ (which somehow establishes the prediction orders, exposed in supplementary data), to finally obtain a result, 'output' (in our case, the clinical report).

To optimize this process, we should work on the following aspects:

• •

  A relatively stable and quality ‘input’: requires structured texts derived from registries. The quality would come from verifying the medical data entered and more granular records that better capture the medical reality. In this regard, work should be done on the ‘integration’ of large national records to enhance efficiency in data entry that allows a better linkage of data to the patient. Interoperability between registries would be a desirable goal.

• •

  If we have a stable structured text, we can iteratively improve our ‘prompt’ (algorithm), so we would get an ‘output’ that was increasingly similar to the sought clinical report.

In this regard, we propose different examples in the supplementary data hybridizing information from RETRAUCI and ENVIN (supplementary data). This approach currently presents limitations that we present in supplementary data.