Back

Neura: a specialized large language model solution in neurology

Barrit, S.; Torcida, N.; Mazeraud, A.; Boulogne, S.; Benoit, J.; Carette, T.; Carron, T.; Delsaut, B.; Diab, E.; Kermorvant, H.; Maarouf, A.; Maldonado Slootjes, S.; Redon, S.; Robin, A.; Hadidane, S.; Harlay, V.; Tota, V.; Madec, T.; Niset, A.; El Hadwe, S.; Massager, N.; Lagarde, S.; Carron, R.

2024-02-13 neurology
10.1101/2024.02.11.24302658 medRxiv
Show abstract

Large language models (LLM) ability in natural language processing holds promise for diverse applications, yet their deployment in fields such as neurology faces domain-specific challenges. Hence, we introduce Neura: a scalable, explainable solution to specialize LLM. Blindly evaluated on a select set of five complex clinical cases compared to a cohort of 13 neurologists, Neura achieved normalized scores of 86.17% overall, 85% for differential diagnoses, and 88.24% for final diagnoses (55.11%, 46.15%, and 70.93% for neurologists) with rapid response times of 28.8 and 19 seconds (9 minutes and 37.2 seconds and 8 minutes and 51 seconds for neurologists) while consistently providing relevant, accurately cited information. These findings support the emerging role of LLM-driven applications to articulate human-acquired and integrated data with a vast corpus of knowledge, augmenting human experiential reasoning for clinical and research purposes.

Matching journals

The top 10 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.