Wednesday March 23, 2022
Entry of Medical Prescriptions Using Natural Language on Smartphones
Abstract :
The use of information technology in health care facilities allows for better organization, strengthens procedures, allows for a continuous information tranmission and makes the care process safer. One of the components of a health information system (HIS) is a prescribing assistance software (PAS) that helps to limit adverse drug events (ADEs). When physicians use software for healthcare, most of the data is entered manually into the PAS, which reduces the time spent on care. To overcome this barrier, we propose to provide a natural language interface to PAS so that practitioners can record their prescriptions using speech at the point of care using a smartphone. Such a system would allow practitioners to use a PAS on the go and allow them to get as close as possible to the most natural way of prescribing medications.

The general approach we have adopted is a task-oriented dialogue system coupled with a PAS destined for prescribers. One of the main challenges was to design a complete specialized dialogue system with processing mostly based on deep machine learning methods without available training data. To circumvent this problem, we proposed an iterative method coupling corpus collection, automatic text generation and expert modeling. We present a conversational modeling validated with medical experts and a detailed study of the characteristics of drug prescriptions from a NLP perspective. One of the main components of a dialogue system is focused on the natural language understanding (NLU) process, which is addressed by a slot-filling approach. Our method of data collection and generation allowed us to create a balanced corpus covering the whole prescription semantics we defined. This corpus allowed the initial learning of the NLU model as well as the boostrapping of the dialogue system.  

To validate our approach and to collect realistic oral prescription data, we proposed an experiment with the dual objective of evaluating the system and collecting data. The evaluation, which included 55 people, including 34 medical experts, showed that the performance of NLU was satisfactory with an F-measure of 90% for a task success rate of more than 75% for the medical experts.  These results are comparable to those obtained on the initial corpus, which confirms that the approach adopted during the thesis was valid. In order to promote reproducible research, the aligned oral corpus (speech-transcription-semantics) comprising more than 4 hours of recordings will be released to the community.

This thesis was carried out within the framework of a CIFRE (convention industrielle de formation par la recherche) collaboration between the company Calystene SA and Laboratoire Informatique de Grenoble (LIG).
Mis à jour le 15 March 2022