Around half of patients with ischemic stroke have mild neurological symptoms (1), usually with the expectation that such patients will come back to their pre-stroke activities regardless of the treatment. However, over one-third of mild stroke patients present with some degree of post-stroke disability (PSD) (2–4), which may be the result of inadequate acute treatments, early stroke recurrence, serious complications, or other reasons (1, 5). For the acute treatment of mild stroke patients, the guidelines from the American Heart Association/American Stroke Association (AHA/ASA) (6) distinguish disabling from non-disabling stroke and recommend intravenous (IV) alteplase only for the former. Nonetheless, the more certain, but not definitive, concept of “disabling stroke” is subjective and requires interpretation by individual neurologists. On the other hand, there is a trade-off between the benefits of IV alteplase and the risk of symptomatic intracranial hemorrhage (sICH). Therefore, decisions on how to treat mild stroke patients should be made on an individual basis.

3-month modified Rankin Scale (mRS), a valuable instrument for testing therapeutic interventions (7, 8), was used to assess the levels of PSD (5, 8). For mild stroke patients who only received medical therapy but had PSD, such therapy is not enough. Therefore, mild stroke patients who would be at high risk of PSD if they only received medical therapy should be early identified in emergency contexts, and some aggressive treatments, such as IV alteplase or close monitoring preventing worsening, should be taken in time. Unexpectedly, neurologists' overall accuracy for identifying those patients was staggeringly low (16.9%) (9). Each day that such a problem continues to exist means that uncounted mild stroke patients are being left with preventable disability.

However, none of the previously published risk models which were developed to predict the function outcome after stroke are fit to solve this problem. For example, the Totaled Health Risks in Vascular Events (THRIVE) score and the Houston Intra-Arterial Therapy (HIAT) score assign 0 points for National Institute of Health stroke scale (NIHSS) ≤ 5, losing the predictive power of NIHSS in mild stroke patients (10, 11). NIHSS on admission has been proven to be a strong predictor of PSD (5). Thus, despite convenient clinical applicability, these models cannot accurately identify mild stroke patients at high risk of PSD. Such models remain inadequate.

With the increased clinical data gathered for each patient, modern medical decision-making demands accurate, novel, and prediction-driven decision support. Machine learning (ML) algorithm, as a burgeoning statistical approach, is well-suited for that mission. Numerous studies with a considerable number of patients have shown great potential for ML approaches to predict recurrence (12), swallowing recovery (13), or aphasia (14) in patients with stroke. However, a model based on ML algorithms, focusing on the more debatable area of treating MS, has not yet been established.

Here, our goal was to develop and validate a prediction-driven decision support tool based on ML algorithms, called DAMS (Disability After Mild Stroke), to early identify mild stroke patients who would be at high risk of PSD if they only received medical therapy, and more importantly, to assist neurologists to make individual clinical decisions for mild stroke patients.

In this study, we demonstrated DAMS had the capacity to early identify mild stroke patients who would be at high risk of PSD if they only received medical therapy, achieving an optimal performance compared with our other ML models and previous scoring systems (THRIVE and HIAT scores). In addition, R-DAMS was developed for more urgent situations. DAMS and R-DAMS were able to generate reliable risk estimates for individuals, relying merely on data that were acquired in an emergency setting, and R-DAMS was able to do this within 4.5 h or less of symptom onset. Hitherto none of the prognosis models for mild stroke patients were developed for the prime objective of providing clinical decision support which targets treatment in the emergency contexts. DAMS and R-DAMS, as prediction-driven clinical decision support tools with this target in mind, are significant because neurologists faced a dilemma about the more debatable area of treating mild stroke: using IV alteplase but with the risk of sICH, or not using IV alteplase but potentially leaving the patient with brain ischemia.

In our study, the use of R-DAMS could offer neurologists effective support in the IV alteplase decision. Whether mild stroke patients will benefit from IV alteplase is still controversial. A meta-analysis reported that mild stroke patients who were treated with IV alteplase had lower odds of PSD even if the incidence of sICH increased slightly (22, 23). However, this research relied on retrospective data. The Potential of rtPA for Ischemic Strokes with Mild Symptoms (PRISMS) trial, which prospectively enrolled mild stroke patients without “clearly disabling” deficits, demonstrated no benefit for IV alteplase in this subgroup of patients (23). This trial defined a more certain, but not definitive, population for which the use of IV alteplase cannot be recommended. In line with the findings of the PRISMS trial, the AHA/ASA guidelines distinguish mild disabling stroke from mild non-disabling stroke and recommend IV alteplase within 3 and 4.5 h only for the former (6). The population in our study was not categorized by whether their initial symptoms were “clearly disabling,” because there are subtle differences in judgments about “clearly disabling” deficits in individual neurologists. In the present study, it should be stated explicitly that for patients who were identified to be at high risk of PSD by DAMS or R-DAMS, medical therapy alone is not enough. Thus, the two models support decision-making in the following ways: First, for mild stroke patients judged to be eligible for IV alteplase by current guidelines, R-DAMS was the best choice. The prediction generated by R-DAMS, paired with neurologists' expertise, enables them to choose the most appropriate candidates for IV alteplase. Second, for patients who are not eligible but are at high risk of PSD according to DAMS, best medical therapy alone with close monitoring may be an appropriate course of action.

On the other hand, we unlocked the potential utility of DAMS in secondary prevention. In a secondary analysis of the Acute Stroke or Transient Ischemic Attack Treated with Aspirin or Ticagrelor and Patient Outcomes (SOCRATES) trial, recurrent cerebrovascular event occurred at a significantly higher rate in patients with PSD than patients without PSD (29.0 vs. 3.7%) (5). Furthermore, as a leading cause of PSD (5, 8), a recurrent cerebrovascular event would do more irreparable harm to the patients at high risk of PSD compared with those at low risk. Therefore, effective prevention of recurrent cerebrovascular event to the patients at high risk of PSD portends a decreased risk of PSD. In the present study, DAMS could help to identify mild stroke patients at high risk of PSD, namely those who would most likely obtain substantial benefits from secondary prevention. For this patient group, a focus on evidence-based treatments for secondary prevention, and a support program to improve achievement of secondary prevention targets (e.g., blood pressure, diabetic control, cholesterol) in the long-term, might significantly reduce PSD.

With the expectation that DAMS and R-DAMS can be integrated into clinical practice, we had to acknowledge that our results represent only one step toward one component of a prediction-driven decision support tool for mild stroke patients. Some other steps need to be considered. Firstly, external validation, using data sets from different centres, should be carried out to duplicate the present results. Secondly, an impact study, quantifying whether application of DAMS and R-DAMS in clinical practice improves neurologists' decision making and subsequent patient outcome, is indispensable (24). Finally, development of simple-to-use software, providing a clear interpretation of the prediction and further treatment/prevention information based on this prediction, is required. The present results are promising but we need to emphasize that much work must be done before completely integrating DAMS and R-DAMS into clinical practice.

In the present study, several predictors of PSD have been discovered. NIHSS is a widespread assessment tool used to quantify the baseline severity in stroke patients. As shown in Figure 4A, even within a narrow range of baseline scores, the strongest feature that contributed to the prediction was NIHSS on admission and the higher the values of NIHSS, the more likely the chance of PSD. Noticeably, although the NIHSS has been widely favored in clinical research, some neurological deficits are measured objectively. For example, one NIHSS item, ataxia, confused hemiplegia and normal function by scoring ataxia as “normal” (0) in patients with hemiplegia (25). In the present study, patients with hemiplegia at admission are more likely to be PSD.

There are some limitations to the present study. Firstly, the mRS used to assess the levels of PSD in our study lacks sufficient detail to describe cognition and mood outcomes. A study published in 2017 in the Stroke journal demonstrates that a considerable number of patients with a good mRS outcome were incapable of socially reintegrating because of cognitive impairment and depression (26). However, the validity and reliability of the mRS was recognized by several clinical researchers (23, 27). Since the mRS is easy to use and interpret, the scale has been a valuable tool for assessing the efficacy of therapeutic interventions till now. Secondly, the lack of external validation in our study hinders the evaluation of external generalizability. As a result, whether DAMS and R-DAMS, which have the selection bias that is inherent in any prediction model, can be used directly in other health institutions is still uncertain. To solve this problem, we provided as much detail as possible about the study cohort (Table 1). This information enables other institutions to judge whether their selected population matches the population here. In addition, the process of model development has been described in a precise fashion in Methods and Supplementary Table 2. Therefore, DAMS and R-DAMS may be transferable to other institutions. Thirdly, recurrent cerebrovascular event, a known predictor of PSD in mild stroke patients, was absent in the process by which DAMS and R-DAMS are developed (5, 8). Our models were initially designed for supporting clinical decision-making in emergency contexts, in which the data of recurrent cerebrovascular event is unavailable.

DAMS and R-DAMS represent one step within a larger process to early identify mild stroke patients who would be at high risk of PSD if they only received medical therapy, by assisting neurologists to make individual clinical decisions for mild stroke patients. Compared with our other ML models and previous scoring systems (THRIVE and HIAT scores), DAMS had a better performance and R-DAMS was able to operate within 4.5 h or less of symptom onset. Future work should build on these findings to transfer DAMS and R-DAMS to different centers.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

Written informed consent was obtained from the individual(s) for the publication of any potentially identifiable images or data included in this article.

XL formed the conception and study design. XC and JZho did the data collection. NC and FJ did the data analysis. SL and DZ did the literature review and model development. XL and SL drafted the manuscript. ZZ, JZha, and JZo made significant revisions and supplied valuable improvement suggestions. The work presented in this paper was carried out in collaboration with all authors. All authors provided approval of the final version. All authors have read and agreed to the published version of the manuscript.

This study received the following financial support: National Natural Science Foundation of China grants 81673511, 81700398, 81970309; and Jiangsu key Research and Development Plan grant BE2017613.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.