1. Introduction
Despite the largely preventable nature of stroke, it remains the second leading cause of death and disability in 2019, which is likely to continue [
1]. Stroke is a complex condition with variable impairments and severity, yet stroke patients are often classified in an overall sense as ‘mild’, ‘moderate’, or ‘severe’ [
2]. This categorization is typically established using the National Institutes of Health Stroke Scale [NIHSS], which is a widely used assessment tool designed to measure neurological impairments in stroke survivors [
2]. The NIHSS is used to screen for neurological impairment across multiple domains, such as consciousness, movement, and language using 11 items and was designed for use primarily in the acute phase after stroke [
3]. The overall score is interpreted as a score of neurological stroke severity according to the following groupings: 0–4 as mild stroke, 5–15 as moderate stroke, 16–20 as moderate to severe stroke, and 21–42 as severe stroke [
4]. This classification often remains with the stroke survivor in the later stages of recovery.
Yet, ‘mild’ stroke survivors often argue the classification of ‘mild’ does not correspond with their daily experiences, as they report depression and difficulties in advanced physical and social activities, leading to a diminished quality of life [
5,
6]. Moreover, studies show that persons with mild stroke struggle to cope with the consequences of stroke, experience difficulties in everyday life [
7], and may undergo persistent disability and difficulty with complex activities [
8]. While it is recognized that the NIHSS is a valid and reliable screening measure, it has been reported that many acute stroke survivors with mild classification by NIHSS could have been overlooked for intensive rehabilitation therapy [
9].
Given the complexity of the physical, psychological, and social burdens associated with stroke, it is important to measure the holistic impact of impairment and recovery following stroke to achieve targeted and personalized care [
10,
11]. This is particularly important for stroke survivors classified as ‘mild’, as survivors of mild stroke may be investigated less due to the assumption that they are expected to regain their premorbid functionality with minimal or no intervention. This is despite evidence that even mild symptoms can impact the ability to perform daily activities and household chores [
12]. Given the physical, social, emotional, and functional burden experienced by persons with mild stroke and considering the common practice of using only the NIHSS to classify stroke severity, further investigation of the latent impairments associated with ‘mild’ stroke survivorship is required, thus leading to the question, ‘Is mild really mild?’
In this study, our objective was to identify various patterns in mild stroke recovery to facilitate tailored and personalized poststroke care. We aim to detect groupings of survivors of mild stroke based on their cognitive, mood, social, and physical abilities, as well as quality of life, analyzing variations in their poststroke experiences beyond the NIHSS-scale-based categorization. For this purpose, we utilized a variety of outcome measures developed to evaluate various aspects of stroke, including motor skills, sensory perception, cognitive function, mood, functional disability, physical capacity, and social interactions among survivors. This more comprehensive profile allows us to better understand the poststroke experiences of survivors and thus improve their quality of life.
These multiple stroke impairment metrics are characterized as multisource and multigranular data, represented as unlabeled data, making these data less amenable to investigation using traditional statistical or supervised machine learning techniques that typically rely on datasets annotated by human experts. Therefore, we utilized a structure-adapting unsupervised machine learning approach, the growing self-organizing map (GSOM) algorithm [
13], to automatically generate profiles of impairments in survivors classified as ‘mild’ by the NIHSS.
We used data from the Stroke Imaging Prevention and Treatment (START) [
14] longitudinal cohort study, which consist of multiple test scores of stroke survivors at three time points [3–7 days; 3 months; 12 months after stroke] in their stroke journey. These time points align with commonly defined phases of recovery: 3–7 days, within the acute phase; 3 months, end of the early subacute phase; and 12 months, within the chronic phase [
15]. Specifically, we processed survivors of mild stroke’ data at these time points to generate GSOM representations at each point in time. By examining the representation captured by the GSOM, we aimed to distinguish unique profiles among ‘mild’ stroke survivors at each time point. These profiles of impairments across multiple domains, including cognition, mood, physical activity and social functioning provide new insights, all without the need for prior knowledge or human annotation.
3. Results
3.1. Demographic and Clinical Characteristics of Stroke Sample
Seventy-three survivors of stroke met the inclusion criteria for mild stroke with available data and were included in this study’s sample. The demographic and background clinical information of the participants is presented in
Table 1. The clinical and functional outcome characteristics of the sample are summarized in
Table 2 for each of the main variables included in the analysis, i.e., NIHSS, MoCA, MADRS, mRS, RAPA, WSAS ,and SIS, at day 3–7, 3 months, and 12 months after stroke.
3.2. Stroke Survivor Clusters Based on the NIH Stroke Scale
The NIHSS consists of 11 items that focus on different neurological aspects such as level of consciousness, horizontal eye movement, visual field test, facial palsy, motor arm, motor leg, sensory, speech, language, and attention. The scores for each question item are aggregated to form the final NIHSS score, which is used for stroke severity classification. However, this aggregated score fails to capture the aspects that are more impaired or not. Therefore, we used the AI framework on the baseline NIHSS item scores, on day 3–7 after stroke, to detect different impairment groupings (annotated regions) that can be derived solely from the NIHSS.
This analysis results in five different profiles pertaining to different impairments, as shown in
Figure 3 and
Table 3. The scores of the participants in such groupings were compared with the rest of the population using t-tests to determine if there was any significant difference. Participants in nonannotated regions did not exhibit patterns that could be differentiated, meaning they had mixed NIHSS item-score attributes that were not significant.
Profile 1 participants (19.2%) showcased mild to moderate somatosensory loss and impairment of motor abilities in the right leg. The impairment scores given to sensory loss (sensory loss mean: profile 1 = 0.571, other participants = 0.068, p < 0.05) and difficulty with motor abilities in the right leg (motor leg right mean: profile 1 = 0.5, other participants = 0.017, p < 0.05) in this profile were significantly higher than those for the other participants. They did not show a remarkable difference in the other attributes.
Profile 2 participants (15.06%) were separated from the other participants due to their increased impairment scores for facial palsy. Profile 2 participants showcased higher impairment scores (facial palsy mean: profile 2 = 1.72, other participants = 0.27, p < 0.05) as the majority were suffering more from partial paralysis of the lower face compared to other survivors of mild stroke.
Profile 3 participants (13.7%) showed increased scores for limb ataxia compared to other participants. All the participants in this profile had ataxia present in either one or two limbs (limb ataxia mean: profile 3 = 1.6, other participants = 0.111, p < 0.05).
Profile 4 participants (8.2%) were differentiated from the rest due to their impairments in speech. All the participants in this profile had mild to moderate aphasia, indicating some obvious loss of fluency or facility of comprehension without significant limitations on ideas expressed (best language mean: profile 4 = 1.33, other participants = 0.07, p < 0.05). Moreover, all the participants had mild to moderate dysarthria, where patients slur at least some words and at worst can be understood with some level of difficulty (dysarthria mean: profile 4 = 1, other participants = 0.313, p < 0.05).
Profile 5 participants (8.2%) showed an increased level of visual impairment compared to the other participants. All the participants in this profile had partial or complete hemianopia, indicating visual impairment (visual field test mean: profile 5 = 1.5, other participants = 0.044, p < 0.05).
This analysis of the NIHSS attributes showcased different subgroupings of impairments among survivors of mild stroke that enabled the creation of another layer of granularity for survivors identified as mild.
3.3. Profiles across Measures
The performance on each of the six domain assessments was mapped for each individual classified as mild. The GSOM generated a latent representation with the six impairment profiles, as shown in
Figure 4 for two individuals. As illustrated, different individuals exhibit impairment in different domains, consistent with their varied poststroke experiences. This provides valuable insights into identifying individual needs that can be considered in the delivery of personalized rehabilitation care.
3.4. Different Profiles of Survivors of Mild Stroke at Different Time Points of Their Recovery Trajectories
The START study obtained measures for participants at day 3–7, 3 months, and 12 months after stroke, which permitted a longitudinal study of stroke impairment and impact. The selected measures, NIHSS, MoCA, MADRS, mRS, RAPA, WSAS, and SIS, were used to assess stroke impairment and impact across several different domains. The GSOM algorithm was applied to participant data at these three time points separately to infer the different profiles of survivors of mild stroke over time. An example is shown in
Figure 5.
3.4.1. Profiling at Day 3–7 after Stroke
At day 3–7 after stroke, only the MADRS and MoCA assessment outcomes were reported in addition to the NIHSS. Based on these data, the AI module separated the participants into two significant profiles indicating clear impairments in these outcomes compared to the other participants, as shown in
Figure 6 and
Table 4. Based on the analysis, 35.6% of the participants reported significant impairments in cognition or depression compared to others.
Profile 1 participants scored lower on the MoCA assessment, indicating lower cognition abilities compared to the other participants (MoCA mean: profile 1 = 19, other participants = 25.77, p < 0.05). They did not show significant impairments in the other domains.
Another grouping exhibited higher levels of depression as they scored higher on the MADRS assessment on day 3–7 after stroke compared to other participants, as shown in Profile 2 (MADRS mean: profile 2 = 16.5, other participants = 3.2, p < 0.05). Participants in the non-annotated regions did not exhibit significant variations in their assessment scores compared with other participants.
Thus, the two identified profiles provided evidence of subgroupings of impairment in cognition or depression on day 3–7 after stroke among stroke survivors who had been classified as mild.
3.4.2. Profiling at 3 Months after Stroke
At 3 months after stroke, the MoCA, MADRS, mRS, RAPA WSAS and SIS scores were reported. Based on these data, the AI module generated three significant profiles indicating different impairments among 35.6% of mild stroke patients, as shown in
Figure 7 and
Table 5.
Participants in profile 1 (9.6%) demonstrated a higher level of depression based on their scores for the MADRS assessment (MADRS mean: profile 1 = 12.71, other participants = 5.47, p < 0.05). Apart from their higher level of depression, they did not exhibit significant impairment in other assessments.
Another notable exemption is profile 2 (12.3%,) where participants recorded higher scores for the mRS assessment (mRS mean: profile 2 = 1.66, other participants = 0.85, p < 0.05), which evaluates the degree of disability following stroke. Participants in this cluster also showed comparatively lower scores on the SIS, which assesses the other dimensions of health-related quality of life: emotion, communication, memory and thinking, and social role function (SIS mean: profile 2 = 64.44, other participants = 84.32, p < 0.05). It could be determined that in contrast to the other groupings among survivors of mild stroke at 3 months after stroke, profile 2 participants displayed more impairment and impact due to their increased disability and poor quality of life.
The AI algorithm separated another group of participants in profile 3 (13.7%), who scored higher on the WSAS assessment (WSAS mean: profile 3 = 15.15, other participants = 3.79, p < 0.05), which indicates low work and social adjustment in daily life. These participants did not show a significant difference in the other assessments.
3.4.3. Profiling at 12 Months after Stroke
At 12 months after stroke, the MoCA, MADRS, mRS, RAPA, WSAS, and SIS scores were used for the profiling. Among participants at 12 months after stroke, 43.9% showed at least one impairment with greater frequency than that of participants with impairment at day 3–7 or at 3 months after stroke. The participants were grouped into identified profiles by the AI algorithm, as shown in
Figure 8 and
Table 6. Four distinct impairment profiles were identified at this time point.
Among the survivors of mild stroke at 12 months after stroke, a group of participants showed lowered cognitive abilities and low engagement in physical activities. These participants, highlighted in profile 1 (11%), scored less on the MoCA assessment (MoCA mean: profile 1 = 21.12, other participants = 26.59, p < 0.05), measuring their cognitive abilities, and on the RAPA assessment (RAPA mean: profile 1 = 2.62, other participants = 4.2, p < 0.05), which evaluated their level of physical activity.
Another group of participants shown as profile 2 (9.6%) demonstrated an increased level of depression compared to other participants as they scored higher on the MADRS assessment for depression (MADRS mean: profile 2 = 9.71, other participants = 4.88, p < 0.05). It is noteworthy that higher levels of depression among this group of participants were seen at all three time points.
Profile 3 (11%) participants were separated from the rest given their increased level of functional disability. This group exhibited higher scores on the mRS assessment, which evaluates disability in stroke survivors for recovery and continued disability. All participants with this profile had disability symptoms, while a few reported moderate disability (mRS mean: profile 3 = 1.5, other participants = 0.75, p < 0.05).
In profile 4 (12.3%), participants indicated a poor quality of life, as indicated by their low scores on the SIS assessment, which evaluates disability and health-related quality of life after stroke. Their SIS assessment outcomes were significantly lower compared to those of the other participants (SIS Mean: profile 4 = 70.55, other participants = 87.67, p < 0.05).
3.5. Capturing Individual Patient Trajectories
Using the GSOM algorithm, we generated patient recovery pathways from the data collected at the day 3–7, 3 month, and 12 month time points, as shown in
Figure 9. The selected subgroup of patients, categorized as ‘mild’ based on their NIHSS overall scores, display varying trajectories despite their initial categorization.
At the 3 month (90-day) period, it can be observed that a cluster of patients present similar patterns of characteristics, with only minor differences in cognitive abilities and work and social adjustment. In comparison, these individuals at day 3-7 show differences in cognitive abilities, as indicated by the MoCA score. Notably, one patient (represented by the orange pathway) exhibited lower cognitive ability and a higher mRS score by 12 months post-stroke. Thus, while the patients exhibited dissimilarities initially, by 3 months their recovery pathways converge, highlighting the dynamic nature of individual recovery trajectories over time.
4. Discussion
Although NIHSS screening is used to provide a measurement of ‘mild’ stroke severity, we examined if individual experiences varied based on other poststroke factors such as their cognition, mood, physical ability, and work and social adjustment. Our findings revealed different groupings (profiles) of survivors of mild stroke based on the GSOM maps. By examining the representation captured by the GSOM, we were able to distinguish distinctive profiles among survivors of ‘mild’ stroke at each time point. These profiles revealed impairments in various domains, including cognition, mood, physical activity, and social functioning, and all without the need for prior knowledge or human annotation. Incorporating such factors from multiple domains adds value to the current NIHSS screening, with potential to better deliver a personalized care plan for survivors of stroke.
The identification of distinct clusters of impairments highlights the need to incorporate a comprehensive assessment of survivors of mild stroke that encompasses evaluations from multiple domains, in addition to the NIHSS screening, to improve personalized care. Acknowledging a range of impairments across various domains can assist clinicians in gaining a better understanding of the diverse clinical profiles associated with ‘mild’ stroke survivorship. This approach can be used to add value to the current neurological screening of stroke, to enhance the quality of life and support planning of home-based rehabilitation programs. This is essential for survivors of mild stroke who experience additional impairments, as they could be omitted from comprehensive rehabilitation care due to the initial screening of stroke severity.
Our findings, using a structure adapting unsupervised machine learning approach, provide new insights into understanding poststroke impairment and recovery for those presenting with ‘mild’ stroke according to the NIHSS during the first week after stroke. First, the investigation of the clustering of impairment across items of the NIHSS revealed that despite the ‘mild’ classification, variations in impairments could be observed. In this analysis of the NIHSS, five such distinct clusters emerged based on patterns of motor disabilities, somatosensory impairment, speech impairment, visual impairment, and facial palsy, highlighting the necessity of providing individualized rehabilitation and care for survivors of mild stroke. We believe that this study is one of the pioneering studies to use an unsupervised machine learning approach to automatically detect different impairment variations in survivors of mild stroke using the assessment outcomes of the NIHSS.
Second, we used additional measures of mood, cognition, and functional outcomes at key recovery time points [
15] to explore variations in the impairment in survivors of mild stroke over time. At day 3–7 after stroke, distinct clusters were defined by the presence of depressive symptoms (based on MADRS) and cognitive impairment (based on MoCA). These findings support the use of mood and cognitive measures at this time as an adjunct to the NIHSS screening.
The granular level of analysis at 3 months after stroke enabled the detection of the three groupings of survivors of mild stroke: a group with higher levels of depression, a group with poorer quality of life coupled with increased disability, and the third group with low work and social adjustment. Among these profiles, special attention should be provided to survivors who have reported poor quality of life and increased disability as this imposes a burden on their daily lives.
At 12 months after stroke, survivors of mild stroke continued to show impairment across different domains despite zero or low scores on the NIHSS. The number of survivors with impairment was higher at 12 months than at previous time points, a potentially unexpected finding [
33]. At this stage, four clusters were detected with significant impairment across multiple domains. One profile reported lower cognition and markedly reduced physical activity. This is consistent with recent reports on deterioration in cognition over time after stroke [
34]. Given the value of physical activity and cognition on quality of life, it is recommended that health professionals continue to monitor and address these outcomes even in those without notable impairment. Another profile showed a higher level of depression. Depression was noted as a key issue in survivors of mild stroke, as all three time points had groups of participants with a significant level of depression based on the MADRS assessment. At 12 months after stroke, profiles 3 and 4 exhibited increased disability and poor quality of life, respectively. Together, these findings emphasize the presence of ongoing impairment and poor functional outcomes across a constellation of domains even at 12 months after stroke in those classified as having a ‘mild’ stroke early after stroke. These impairment profiles at 12 months after stroke highlight the burden carried by survivors of mild stroke across different domains despite their initial classification as mild.
Capturing personalized longitudinal pathways is crucial for tailoring treatment plans to individual patient needs. By monitoring recovery trajectories over time, healthcare providers can identify specific patterns and variations in patient progress that might otherwise be unnoticed. This detailed insight allows for the customization of interventions based on individual recovery rates, cognitive abilities, and overall health status. Personalizing treatment plans based on longitudinal data ensures that each patient receives the most effective care, improving outcomes and optimizing resource utilization. Moreover, understanding these unique pathways can lead to better-informed clinical decisions, ultimately enhancing the quality of life of patients through more precise and responsive healthcare strategies.
The presence of impairment across multiple domains advocates for survivors of stroke classified as ‘mild’ in the first week after stroke being closely monitored, at least over the first year after stroke, and/or be offered bursts of rehabilitation to prevent or address these ongoing impairments. Our findings also provide strong evidence supporting the voice of people classified as having a mild stroke that ‘mild’ is not really mild, based on their lived experience.
Given the fact that stroke affects the physical, cognitive and mood functionality of a person, it is imperative to identify and understand these complexities [
35]. Wide variability in quality-of-life ratings (0.45 to 0.95 on a scale from 0 to 1) was reported even in those with mild stroke [
35]. Yet, it was established that vague measures aimed at determining the quality of life following stroke impede clinician decision making [
36] as survivors of mild stroke report an abundance of issues associated with return to meaningful activities and life satisfaction [
7,
8]. These findings, together with the current findings, suggest that the NIHSS screening alone does not adequately capture the underlying reality of survivors of stroke. Rather, it suggests the value of a profile of outcomes to provide a more meaningful and comprehensive view of stroke survivorship and quality of life.
Several implications for clinicians arise from this study. First, we provide evidence to showcase different profiles of impairment that exist among survivors of ‘mild’ stroke at different times in the first year after stroke. The fusion of data from multiple assessments enabled the generation of an overview for each person, which is otherwise challenging to assess using conventional means. This new approach permitted the illustration of different profiles of stroke survivors despite the single ‘mild’ classification by the NIHSS. The evidence presented in this paper relating to various groupings of survivors of mild stroke confirms that the stroke severity classification should not rely only on neurological functions but would benefit from incorporating cognition, mood, functional disability, physical, and social activity measures and self-perceived impact. Based on our results, we propose that an optimal approach for assessing stroke recovery would integrate multiple existing scales to provide a more comprehensive view of a patient’s recovery process. While each scale mentioned in our manuscript—including the Montreal Cognitive Assessment (MoCA), Montgomery–Åsberg Depression Rating Scale (MADRS), modified Rankin scale (mRS), Rapid Assessment of Physical Activity (RAPA), Work and Social Adjustment Scale (WSAS), and Stroke Impact Scale (SIS)—effectively measures specific aspects of poststroke impairment, none captures the entire recovery spectrum alone.
Our findings suggest that combining these scales allows for a more nuanced profiling of stroke survivors. Furthermore, our findings indicate that at different time points, different scales were distinctive in the profiling, e.g., MoCA and MADRS on day 3–7; MADRS, mRS/SIS, and WSAS at 3 months; and MoCA/RAPA, MADRS, mRS, and SIS at 12 months. By clustering and analyzing the combined outputs of these diverse scales, we can better identify patient profiles and tailor interventions accordingly. New insights from and use of the AI clustering approach now makes profiling across multiple scales feasible for clinicians. Alternatively, this integrated approach could lead to the development of a new, comprehensive stroke assessment tool that more holistically evaluates cognitive, emotional, physical, and social recovery aspects. Such a tool could significantly enhance personalized care plans and improve overall patient outcomes. Our approach enables widening the scope of the monitoring of stroke survivors and demonstrates the value of incorporating multiple domains in the characterization of survivors of stroke. This is significant in the medium to longer term, when survivors of stroke continue to experience impairments and impact despite being classified as ‘mild’.
Second, the identification of distinct profiles of impairment and impact at different times enables the provision of personalized and targeted care and rehabilitation to survivors of stroke focusing on the domain and profile of impairment. For example, the identification of cognitive impairment and depression in survivors of ‘mild’ stroke may initiate treatments related to emotional health, cognition, and quality of life. This could promote intervention therapies, as the early detection of survivors of stroke with similar levels of depression could facilitate counselling and evidence-based care at early stages. Such need-based care and precautions would in turn enrich the poststroke quality of life of survivors of ‘mild’ stroke, thereby improving the current rehabilitation and person-centered care.
Third, we suggest the plausibility of integrating AI-enabled insights for decision making and designing strategies for rehabilitation that are associated with improving function and the quality of life in survivors of stroke. Using the framework presented in this paper, clinicians can input data related to different stroke assessments to visualize distinct subgroupings of stroke survivors, as shown in
Appendix A. The framework is scalable to accommodate data from many patients and operate on a larger scale. While this serves as a cost-effective decision-making platform, it also categorizes stroke survivors based on the similarity of their impairments, permitting clinicians and therapists to strategically design treatment and rehabilitation programs for survivors of stroke who have similar disabilities. Using this approach in healthcare institutions to analyze poststroke patient data has practical implications: The interactive visualization tool puts the approach and resources into the hands of clinicians. It can help identify where an individual is positioned in relation to the clustering of impairments that may impact the recovery trajectory, allowing therapists to personalize care plans more effectively.
As limitations of this study, we acknowledge that the portion of participants with missing data could be improved. This occurred primarily due to patients not performing all the tests planned in the study. Furthermore, information from survivors categorized as having moderate and severe severity could also be used as additional information for comparison purposes.