Artificial intelligence in pediatric stroke: An ally in screening?

OBJECTIVE Evaluate the ability of ChatGPT for diagnostic guidance in patients with stroke suspicion and its ability to predict the likelihood of stroke. Additionally, to analyze its use as a scale calculator (Glasgow Coma Scale[GCS], Pediatric National Institute of Health Stroke Scale[PEDNIHSS]).

METHODS An observational, descriptive, retrospective, single-centre study was conducted in a Paediatric Emergency Department (PED). Standardised templates were created from the medical records of patients in whom the PSC had been activated and had been attended at the PED. These templates were introduced into independent conversations with ChatGPT. ChatGPT was instructed to: perform diagnostic guidance, predict stroke probability (in a 1-5 scale), calculate GCS and PEDNIHSS. Results were compared with the clinical diagnosis (stroke or stroke mimic) after assessment at the PED.

RESULTS 116 PSC activation case templates were analysed. ChatGPT's diagnostic guidance had weak concordance (k = 0.291), sensitivity 90.9%(95%CI78.9-100) and specificity 57.4%(95%CI47.5-67.4). The AUC of ChatGPT regarding the stroke probability scale, was 0.796 (95%CI:0.714-0.878). The optimal value for stroke detection on the stroke probability scale was 4, with sensitivity 88.24% (95%CI:77.41-99.07), and NPV 93.75% (95%CI:87.82-99.68). The concordance for the calculation of scales (GCS and PedNIHHS) with respect to the real value was weak (k = 0.227, k = 0.261 respectively).

CONCLUSION ChatGPT does not seem useful for diagnostic guidance nor scale calculation in patients with stroke suspicion. However, scores <4 on the stroke probability scale could help to rule out stroke, hence Chat-GPT could be a support tool in PSC activations.