UNED has collaborated in the creation of an artificial intelligence system capable of automatically transforming medical reports into standardized clinical codes, an essential step to structure health information and facilitate its exploitation and use in research.
As detailed by the university, the project proposes an architecture that converts medical texts into codes of the International Classification of Diseases (CIE) through a three-stage workflow: entity recognition, supervised classification, and semantic similarity analysis. Tested on Spanish and English corpora, the system has obtained competitive results and an average improvement of 3.42% in the F1 metric compared to previous approaches.
Clinical coding makes it possible to translate the content of reports -diagnoses, symptoms, procedures or history- into a common language based on standardized codes. In this way, professionals from different centers can record and consult data homogeneously, which simplifies health management and the subsequent analysis of large volumes of information.
Despite this, specialists in the sector recall that this task continues to be carried out largely manually, with a high consumption of time and human resources. "Systems that automate the process provide a lot of value, as they convert a very tedious and time-consuming task into a fast and efficient process, which in turn allows more time to be dedicated to investigating and analyzing information," explains Alicia Ramírez, researcher of the NLP&IR group of the Department of Languages and Computer Systems of UNED and participant in the development of the system.
The proposal also integrates two relevant contributions. On the one hand, it incorporates an unsupervised phase that allows locating codes that the model had not found during training, expanding its adaptability to real scenarios. On the other hand, it can interpret complex references present in the clinical text, such as fragmented or superimposed mentions that must be understood jointly to assign the appropriate code.
An AI capable of justifying its decisions
One of the most novel elements of the work is that the system does not act as a black box. In addition to generating a proposed coding, it highlights the specific fragments of the medical report that support each prediction. "The presented system, in addition to predicting ICD-10 codes, returns the parts of the text that justify these predictions," indicates the researcher. This allows healthcare personnel to understand why a certain code has been chosen and to validate the result more quickly and transparently.
To validate its behavior, the system underwent tests with corpora in Spanish and English. Although the models were trained separately for each dataset, the objective was to demonstrate that the methodology maintains its performance in diverse contexts and is not limited to a single type of clinical documentation.
According to Ramírez, the improvement achieved - 3.42% in F1 - is especially significant in a task that is very demanding from a computational point of view. The expert recalls that automatic clinical coding operates with highly specialized medical language and with more than 100,000 possible codes, so that apparently modest advances have a notable effect on the system's accuracy and coverage.
The next step in the research will be to build a functional demo that allows the introduction of clinical texts and automatically displays both the suggested codes and the fragments that justify each decision. If the results are satisfactory, this line of work could open the door to future applications in care or research settings.
The development has been carried out thanks to the UNED infrastructure and the specialized knowledge of the NLP&IR group in natural language processing applied to the biomedical field, a consolidated line of work within the Department of Computer Languages and Systems.