The Prognosis of Ischaemic Stroke and Intracranial Haemorrhage Imaging-Based Subtypes; Matched Cases and Controls in Healthcare Systems Data

Background and aims
Outcome after stroke varies by stroke type and stroke subtype, but existing studies of stroke subtypes are limited by cohort size or outcome data availability. We used natural language processing (NLP) of brain imaging reports to estimate risk of death and other outcomes after stroke subtypes in a nationwide dataset.
Methods

We applied our validated NLP algorithm, EdiE-R, to all CT and MRI head scan reports in Scotland from 2010-2018. We linked the resulting labels to hospital admissions, prescriptions and death data. We used NLP to categorise people with non-specific ICD-10 stroke codes, and to identify subtypes of ischaemic stroke and intracerebral haemorrhage (ICH) by location (cortical/lobar and deep). We compared stroke recurrence, myocardial infarction, cancer, dementia, epilepsy and death to people without stroke using a matched case-cohort design.

Results
We identified people with ischaemic stroke (n=51339;14103 cortical and 12616 deep), ICH (n=7182; 1465 lobar and 1814 deep), subarachnoid haemorrhage (n=2984) or subdural haemorrhage (n=2714). Outcome event rates varied between stroke subtypes. Key findings included higher hazard ratio for early mortality for ICH (23.06 (20.79-25.58)) than ischaemic stroke vs controls (10.89 (10.02-11.83)), lower hazard ratio of readmission with stroke for deep compared to lobar ICH (0.59 (0.39-0.87)) (Figure 1) and higher hazard ratio of long-term dementia for lobar ICH (3.5 (2.3-5.3)) than for deep ICH (2.3 (1.5-3.5)) or ischaemic stroke vs controls (1.97 (1.84-2.11)).

Conclusions
NLP allows subtyping of stroke on a national scale, providing real-world information on outcomes after stroke. The method could be replicated in other datasets across Europe.