STOP-stroke: STroke Outcome Prediction in the Acute Treatment Setting - a Prospective, Single-center, Observational Study

To avoid irreversible brain damage in acute stroke, neurologists must make treatment decisions under immense time-pressure. In current clinical practice, neurologists decide using visual inspection of brain scans and established clinical parameters. Despite the large quantity of data, statistical or machine learning models have not yet reached clinical practice to guide decision-making. This is in large part because doctors cannot assess the trustworthiness of these models, and because current models cannot handle multimodal input data. Since the field of acute stroke treatments is constantly evolving, there is an urgent need to improve our understanding of the factors determining stroke patient outcome and response to treatment. We are convinced that the project will provide new insights into stroke outcome prediction and help to integrate data from machine learning algorithms into clinical routine.

Neurologist's prediction of stroke outcome When a patient arrives at the stroke unit, the treating team continuously integrates information as to what the best therapeutic options or risk of complications may be. To predict patient outcomes, neurologists rely on structured tabular data, such as age, sex and blood pressure, and unstructured data, such as medical image data. Current prediction of therapeutic success in stroke patients outside the therapeutic time window is often based on brain images (Joundi et al. Neurology 97:S68-S78). Diffusion and perfusion weighted imaging (DWI/ PWI) are used to identify the infarct core and hypoperfused region, resulting in an estimate of the tissue at risk referred to as "mismatch"(Heiss WD et al. Int J Stroke 14:351-8). Large mismatch and small infarct core are considered surrogate markers for treatment success, which is why clinical trials have often pre-selected such patients (Albers GW et al. N Engl J Med 378:708-18). However, recent studies question the validity of such a pre-selection. For instance, novel, pooled data analyses indicate that patients with large infarct core benefit from endovascular treatment as well (Campbell BCW el a. Lancet Neurol 18:46-55). Moreover, studies with the pre-selected patients do not allow drawing any conclusions about the effect of treatment in patients with a lack of mismatch or large cores. It becomes increasingly clear that patients with a large core or lack of mismatch could benefit from endovascular thrombectomy treatment as well (Karamchandani RR et al. J Stroke Cerebrovasc Dis 31:106548). We expect that a refined modelling approach and the integration of imaging and clinical parameters will yield better prognostic factors and more reliable outcome predictions.

Computer aided prediction of stroke outcome For computer aided outcome prediction in stroke patients, scores were developed based on a few preselected tabular features, such as age and sex, and an underlying logistic regression model. While these methods slightly outperform the score-based methods, they lack interpretability. Therefore, currently there are no computer aided outcome prediction tools used in clinical routine.

Previous preparatory work In a previously conducted recent project, weanalyzed imaging features from DWI and perfusion imaging data such as TTP, CBF,CBV and TMAX in a group of stroke patients with LVO, all treated with MT at the InselSpital Berne. We tested if the generally accepted mismatch concept, derived from routine clinical software, indeed enhances outcome prediction for individual patients. We demonstrated that neither the core nor the mismatch volume significantly improved prediction of outcome when used in addition to clinical parameters (Hamann J et al. Eur J Neurol 28:1234-43). To implement a stroke outcome prediction model that is trustworthy for clinicians, we developed deep learning (DL) based models that can not only integrate structured and unstructured data, but also yield interpretable parameter estimates for clinical parameters like odds ratios (https://www.sciencedirect.com/science/article/abs/pii/S003132032100443X). We applied them to data from ischemic stroke patients treated at the University Hospital Berne (https://arxiv.org/abs/2206.13302) and investigated if our models can compete with experienced neurologists by performing a blinded prediction challenge, where five neurologists and our trained DL model were provided with the same structured clinical data of stroke patients. In this experiment, our model tended to show a better performance, particularly when image data were provided (Herzog L et al. Stroke. 2023 Jul;54(7):1761-1769).

Accordingly, the present STOP-stroke study aims to improve outcome prediction early after stroke by assessing both the clinician's outcome estimation as well as our trained DL model. Thereby, we try to implement a new prognostication tool, which can be helpful in supporting treatment decision and individualized patient care.