INTRODUCTION

Stroke is the leading cause of acquired brain damage and disability, with an increasing incidence in developed countries¹. Emergency care for a patient with a suspected stroke centers around the “time is brain”, principle due to the importance of quickly identifying patients with an ischemic stroke to prevent potentially irreversible damage to brain tissue. This scientific statement recognizes not only the importance of the nurse’s role in recognition of typical and atypical symptoms of stroke and prompt activation of the stroke response system, but also in all the stages of the evolution of the stroke, including the prognosis^2,3.

Ischemic stroke caused by arterial occlusion is responsible for most strokes (more than 70%). Given that it is a time-dependent situation, correct management should include rapid and accurate triage and diagnosis^1,4, and preventing not-so-frequent presentations from escaping it. Thus, it will be possible to immediately activate the stroke code, and achieve effective reperfusion of the ischemic brain area with definitive treatment such as intravenous fibrinolysis and/or endovascular mechanical thrombectomy. While the particular type of stroke may be suspected before arriving at the hospital, or on the triage set by the nurse, it cannot be confirmed without diagnostic imaging. Rapid processes are required in the hospital, and a team approach, to reduce the time from hospital admission to definitive treatment (diagnosis-needle time)^5,6.

In some countries, the existence of specific stroke nurse has been pivotal in the early management of stroke patients, in the prognosis, and in the post-stroke neurological status at the moment of the discharge from the hospital. The “D’s of Stroke Care” are identified as significant steps in the management of stroke where nurses facilitate time-sensitive care are: Detection and education programs (nurses are key to community education programs that teach risk factor awareness and recognition of early stroke warning signs). Dispatch (nurses in the triage set are often the first point of contact for someone having a stroke). Delivery (comprehensive assessment of a suspected stroke patient). Door (nurses are often the link between the prehospital and the emergency department teams). Data (in order to determine processes and options). Decision (the goal is to begin stroke treatment within 30 minutes after arrival at the hospital). Drug/Device (treatment with clot dissolving medications, as well as appropriate patient monitoring during administering of the medications). Disposition (nursing interventions during transitions when a patient leaves the emergency department can successfully improve stroke outcomes)^7,8.

A wide variety of factors influence stroke prognosis, including age, severity, mechanism, infarct location, previous illness and treatments, clinical findings, and related complications. In addition, interventions such as thrombolysis, mechanical thrombectomy, and stroke unit care, can play a major role in the outcome of ischemic stroke. 90% of acute ischemic strokes can be attributed to modifiable risk factors, and approximately 75% are due to hypertension, lipid metabolism alterations, diabetes, arrhythmias, and behavioral factors such as smoking, poor diet, and poor physical activity. Achieving control of risk factors could prevent more than three-quarters of the global stroke burden⁹. It is well known that several clinical indicators, such us the National Institute of Health Stroke Scale (NIHSS) strongly predicts outcome after stroke¹⁰. Thus, the knowledge of the important factors that affect prognosis is necessary to make a reasonable prediction for individual patients, to provide a rational approach to patient management, and to help the patient and family understand the course of the disease.

Currently, the large amount of information stored in databases about multiple patients who have suffered this pathology can be processed with machine learning techniques in order to automatically detect the most relevant factors and predict the evolution of future patients. Machine learning is a process used to obtain knowledge from large sets of raw data. It involves the application of algorithms to extract patterns from data. These fall into two main categories, supervised and unsupervised learning algorithms. Techniques in the former category are used to induce predictive models and those in the latter to discover similarities or correlations that are hidden in huge volumes of data. There are many applications domains where large databases can be exploited with machine learning algorithms. One of them is the health area since it is a rich data source, including electronic medical records, administrative reports, and medical imaging among others^11,12. There are numerous studies in the literature, including from our group^{13)(14)(15)(16}, in which different machine learning algorithms have been used for various purposes such as the automation of medical diagnosis, and prediction of mortality or treatment outcomes.

Discharge to home from rehabilitation marks an important milestone in the life of a stroke survivor. A lot of stroke survivors are being discharged home to cope with stroke-related disabilities. Although numerous studies have looked at predictors of discharge to home, and the factors that influence the functional deficits of stroke patients, in particular the multiple factors that influence the neurological status of the post-stroke patient, no machine learning techniques have been used in this area for this purpose. There is some related work but with a different objective, such as the use of these algorithms to classify the time elapsed since stroke onset based on image features¹⁷.

In a previous paper in this journal¹⁸ we have highlighted the role of nursing in triage and prediction of the stroke, based upon the scales performed by nurse in the triage area of the hospitals. In the present study we strongly amplify these previous results, with the objective of obtaining a more accurately information about the early prediction of neurological prognosis, in particular the functional independence immediately after discharge from the hospital. Thus, a combined clinical analysis and machine learning techniques of the possible associations in rules of factors that influence the neurological status of the post-stroke patients at the moment of the discharge from the hospital has been carried out.

MATERIAL AND METHODS

A total of 2080 medical records of patients with confirmed acute stroke were analyzed. The area of influence of this study was the Salamanca University Hospital and the 112 emergency service of Salamanca during the years from 2016 to 2019. This area of influence includes not only the entire province of Salamanca but also the neighboring provinces of Ávila and Zamora, which represents a registered population over 664.000 people. We must point out that Salamanca has a very high unregistered population, due to the influence that the University of Salamanca has on this data. In fact, there is an unregistered population of more than 50,000 university students, although, due to their age, the repercussion in a study of stroke, a pathology with a great relationship with age, is of minor importance, but must be considered.

We have only considered ischemic strokes in this study, ruling out hemorrhagic ones. We have used only those cases in which the stroke code was activated. Thus, the final number of medical records analyzed in the four years was 1572:

The information of the patients was obtained from the medical records under the strict control and supervision of the Documentation Service of the University Hospital of Salamanca. Data management was carried out under the rules of the Ethics Committee for the use of the information collected in the patients' medical records.

The total of parameters analyzed of the medical records have been divided into 11 groups:

The present work focuses on inducing patterns and rules from the previously described variables in relation to the status at the moment of the hospital discharge. All the data were analyzed in order to:

RESULTS

Clinical descriptive data (clinical atlas)

A minor predominance of male patients (871, 55,40%), when compared to females (701, 44.60%) was found in our study. It should be noted that as age increases, the percentage of women with respect to men increases. Thus, in those over 80 years of age, the number of women is greater than that of men (60.50% versus 39.50%). On considering age, was a subdivision into 3 age ranges, was performed. 70 patients (4.46%) were placed in the group under 50 years of age. The group between 50 and 80 years old presented a total of 908 (57.76%), with 594 patients in the range of older than 80 years (37.78%).

Regarding the relationship between variables, gender and age have been selected (Image 1), and their relationship with the state at discharge, as the latter is the most important variable in this study (Images 2and 3).

Image 1: Relationship between variables (gender and age). 

Image 2: Relationship between gender variables and status at discharge (OH: another hospital, DM: major deficits, DN: minor deficits, RAD: recovery ad integrum, EX: death). 

Image 3: Relationship between variables age and status at discharge (OH: another hospital, DM: major deficits, DN: minor deficits, RAD: recovery ad integrum, EX: death). 

Analyzing previous diseases (see Image 4), hypertension has been the parameter with the greatest presence prior to the incident, with a total of 583 patients which represents 37.08%. 391 patients had permanent, paroxysmal, or de novo atrial fibrillation at the time of ischemic stroke diagnosis. An additional 128 patients presented some other type of arrhythmia, highlighting atrial flutter and sinus arrhythmia. Thus, the total number of patients with some type of previous arrhythmia was 519, which represents 33.01% of the total. A total of 452 patients presented some type of lipid metabolism alterations, reaching 28.75%.

Image 4: Previous diseases. 

Diabetes was present in 443 patients. Comparing to groups, type 2 had a higher expression (286, 18.2%) than type 1 (157, 9.91%). Of the total of 1,572 patients analyzed, 223 had non-ischemic heart disease, and 121 had previous ischemic heart disease. Thus, 344 patients had some type of heart disease at the time of diagnosis of ischemic stroke, which represents 21.88% of the total number of patients. Lipid metabolism alterations accounted for a total of 452 patients, which represents 28.75% of the records. All alterations in lipid metabolism have been considered as such, regardless of whether the altered parameters were one or several (cholesterol, high density lipoprotein or low density lipoprotein…).

A very interesting aspect has been the associations of previous pathologies observed in many of the patients who suffered stroke. They have occurred significantly among the five most frequent pathologies. Considered in pairs, the association of hypertension with lipid metabolism alterations has been present in a total of 326 patients, which represents 20.74%. The association of hypertension with arrhythmia is also very frequent, particularly with atrial fibrillation (282. 17.93%) and, to a lesser extent, hypertension with diabetes mellitus (236. 13.47%). An important group of patients presented triplet association, the most frequent being formed by hypertension, alterations of lipid metabolism and arrhythmia, singularly atrial fibrillation (201. 12.79%). Another triplet, particularly frequent, was formed by hypertension, heart disease and diabetes mellitus (172.10.94%).

By analyzing previous treatments (Image 5) 525 patients (33.39%) had prescribed diuretics prior to the stroke, both loop, thiazide, and potassium-sparing. The most frequent of all, far ahead of the rest, was furosemide. A total of 223 patients (14.16%) were receiving treatment with digoxin.

Image 5: Previous treatments. 

338 patients (21.51%) received treatment with ACEIs. 223 patients (14.18%) were in antihypertensive treatment with ARBs. 208 patients (13.23%) were treated with beta-blockers, the most frequent of all being bisoprolol.

259 patients (16.47%) were receiving treatment with acenocoumarol at the time of the stroke. 207 (13.16%) were taking new-type anticoagulants (rivaroxaban, dabigatran, apixaban, and edoxaban).

377 patients (23.98%) were treated with different statins. The most used were atorvastatin and simvastatin. 208 patients (13.23%) were taking betablockers (most of them bisoprolol).

The two most affected arterial territories in this study were, far above the rest, the middle cerebral arteries (image 6) with a total of 751 patients, which represents 47.77% of all stroke codes. Considered separately, 379 (24.11%) correspond to the right middle cerebral artery and 372 (23.66%) to the left middle cerebral artery. A total of 166 patients (10.56%) presented occlusion of the basilar artery, 88 (5.60%) of the right vertebral artery, 69 (4.39%) of the trunk of the right carotid artery, and 67 (4.26%) of the right posterior cerebral artery.

Image 6: Most affected arterial territories. 

The rest of the cerebral arteries presented lower and quite similar occlusion Figures. The left anterior cerebral artery accounted for a total of 59 cases (3.75%), as well as tandem lesions. Similar Figures for the left posterior cerebral artery (55, 3.50%), the right anterior cerebral artery (52, 3.31%) and the left carotid trunk (50, 3.18%). The three arteries with the lowest frequency of occlusion were the left vertebral artery with 46 cases (2.93%), the left internal carotid artery with 44 (2.80%), and the right internal carotid artery with 32 (2.04%).

A total of 287 patients (18.26%) received fibrinolysis (alteplase). 330 (20.99%) were proposed for thrombectomy, as they met the requirements for it. However, finally just 255 received the mechanical endovascular treatment (thrombectomy) (Image 7).

Image 7:  Fibrinolysis and thrombectomy. 

Analyzing, the cardinal presentations, the clinical data with the greatest expression was language disorders with a total of 285 patients (18.13%), followed by the group of decreased level of consciousness, with 249 cases (15.84%). Motor problems of the extremities, partial (231, 14.69%) and total (189, 12.02%) ranked fourth and fifth, respectively. Behavioral alterations were present in 7.44% of the total cases (117 patients, Headache, was confirmed as the onset in 56 cases (3.56%). Visual disturbances showed a very low expression with a total of 31 cases (1.97%).

Considering the NIHSS scale, the most important number of patients (474, 30.15%) was situated in the range between 12-15 points. 435 (27,67%) patients were in the range of 16-20 points. 387 patients (24.62%) were situated between 7 and 11 points, while 186 (11.83) were in the area with the highest score above 20 points. With less than 7 points, a total of 36 patients (2.29%) were observed. Analyzing the MRS scale, a total of 808 (51.40%) showed 0 points, indicating that their previous baseline status allowed them to be valid for all possible activities of daily life. The next two groups with the largest number of patients were those that presented a score of 1 (379, 24.11%) or 2 points (11.13%).157 patients (9.99%) presented at the time of stroke diagnosis a total of 3 points, and 38 patients (2.417%) added 4 points. No cases of patients with 5 or 6 points were observed.

A total of 706 patients (44.91%) were discharged with major deficits (Image 8), which seriously hindered their activities of daily living, requiring a permanent caregiver. 674 patients (42.87%) presented minor deficits at the time of discharge. They were valid for most of the activities of daily life, but in some of them, they presented the need for specific help. Of the total of 1572 patients considered in this study, 96 (6.11%) were discharged with recovery to normality, ad integrum, or minimal symptoms that did not impede or hinder, without external help, activities of daily living. 27 patients (1.72%) were transferred to another hospital outside the Autonomous Community of Castilla y León, and a total of 69 patients (4.39%) died before being discharged.

Image 8: Neurological status at the moment of the discharge. 

DISCUSSION

Machine learning, simply explained, is a process used to extract useful information that is hidden in large volumes of data. Association patterns obtained through machine learning can be used to make predictions or simply to find hidden relationships in the clinical information. The data collected in this study are the first processed with machine learning techniques, to the best of our knowledge, on the stroke code worldwide. Our study reveals data association patterns using this approach, some novel and others confirming previously existing patterns with other statistical and data management techniques^9,10.

Clinical decision making is essential in stroke, not only at the moment of the triage by the nurse, but also in the previous detection and control of risk factors⁴. Today, it is well known that the main risk factors for stroke include hypertension, diabetes mellitus, lipid metabolism alterations, arrhythmias (especially atrial fibrillation), and cardiovascular diseases^3,9. These five diseases are considered “the five (not four!) horsemen of stroke apocalypse”. The knowledge of them, and the importance of the most dangerous associations of them in the stroke prognosis are a clue instrument for the nurse.

In many application domains, such as medicine, the datasets used to drive classification models often have an uneven distribution of instances of each class. This causes the results being based mainly on the majority class, obtaining few patterns from the minority class and, in prediction problems, a significantly lower precision for this class. Often the minority class is the most interesting in terms of the application domain being studied, so the problem is compounded in these cases. Our own team, in previous works and publications analyzing different pathologies and medical situations, has detected these problems^{13)(14)(15)(16}.

In order to reduce the dimensionality of the data, one of the problems of our analysis, we have used feature selection methods in our study, which has allowed us to obtain association rules with higher support and confidence values. Likewise, as shown in the results section, the application of this type of method has resulted in an increase in the reliability of the prediction models. To deal with the problem of data imbalance, resampling strategies have been used. Specifically, in our study we have applied the SMOTE technique, which is the one that has provided us with the best results, as in previous studies^{13)(14)(15)(16}. With the resampling of the data, we have managed to obtain a greater number of association rules that exceed the support and confidence thresholds. As we have showed in the results, machine learning clearly shows that age, hypertension, atrial fibrillation, and diabetes mellitus are as the most important factors in determining the neurological status at the moment of discharge from the hospital.

Today, it is well known that hypertension is a powerful determinant of ischemic stroke risk, with clear previous data showing that controlling blood pressure levels <140/90 mmHg reduces the risk of stroke. Evidence of benefit is weaker for lower blood pressure targets achieved with aggressive reduction, especially in older patients^22,23.

Diabetes mellitus is an important risk factor for stroke²⁴. It causes pathological changes in the blood vessels and can lead to a stroke if the brain vessels are directly affected. In addition, mortality is higher and post-stroke outcomes are poorer in patients with uncontrolled glucose levels. Control of diabetes and other associated risk factors are effective ways to prevent strokes, as well as their recurrence^24,25. As it is a disease mainly associated with lifestyle, patients with diabetes mellitus 2 usually have additional risk factors for stroke, such as obesity, hypertension, and dyslipidemia, which multiplies the vascular risk in these patients²⁴. Our results show an association of hypertension, heart disease, and diabetes mellitus in a significant number of patients (172, 10.94%).

There is a complex relationship between normal serum lipid levels and stroke risk. When stroke is subtyped into ischemic and hemorrhagic, contrasting relationships are found between them. High levels of total cholesterol are associated with an increased risk of ischemic stroke, while low levels are associated with an increased risk of intracerebral hemorrhage²⁶. Within ischemic stroke, the subtype of atherothrombotic and occlusive disease of the large arteries has been more closely related to higher levels of total cholesterol. In our study, lipied alterations was one of the pathologies with the greatest presence, reaching a total of 452 patients, confirming it as a predictive clinical history of ischemic stroke, and forming part of the five horsemen of the stroke apocalypse.

Arrhythmias in general, and atrial fibrillation in particular, significantly increase the risk of ischemic stroke. Oral anticoagulation reduces the risk of stroke^27,28. It is estimated that about 40 million people have atrial fibrillation in developed countries. During the last decades, we have learned that this dysrhythmia originates from the interaction between genetic predisposition, ectopic electrical activity, and abnormal atrial tissue substrate. Atrial fibrillation and stroke have been associated in rigorous studies, indicating a true association rather than a false finding^27,28.

Five rules were obtained with a confidence above 0.85, highlighting the association of an NIHSS in the interval 11-16, and the performance of thrombectomy and fibrinolysis, with a confidence of 0.93, for a high ad integrum recovery. These results would be comparable to previous series with descriptive studies and traditional statistics, in which the NIHSS are not very high, and the performance of revascularization techniques, particularly when they are successful, suppose an evident improvement in the functional status of the patient at discharge. and in the following months. There is therefore no doubt, and it is an irrefuTable fact in 2022, that the main objective of advanced treatment of acute ischemic stroke, and the end of the activation of the stroke code, is the early revascularization of the cerebral ischemic area and the limitation of neuronal damage. Mechanical thrombectomy doubles the odds of a better functional outcome compared to standard therapy alone, with no significant difference in mortality or risk of parenchymal hemorrhage at 90 days. Mechanical endovascular treatment should not be avoided on the basis of age alone, and patients older than 80 years may also benefit from it. Although the neurobiological rationale underlying these subgroups remains to be explored, factors that may contribute to different recovery rates include the amount of perilesional edema, as well as remote effects on distant but structurally and functionally connected brain regions, due to injury^3,29.

In our study, a total of 706 patients were discharged with major deficits (44.91%), 674 with minor deficits (42.87%), 96 with ad integrum recovery (6.11%) and 69 (4.39 %) were exitus at some point in the entire process since the stroke code was activated. 27 patients (1.72%), for different reasons, normally return to their reference hospital, were discharged before fulfilling the requirements for it. As we have already commented in the results section, the group with the greatest deficits normally presented the highest NIHSS at the time of diagnosis, together with the fact that it was not possible to apply fibrinolysis or thrombectomy, either one of them, or both techniques. Additionally, and logically from the clinical point of view, discharges with recovery ad integrum used to correspond to low NIHSS together with the application of at least one of the two effective treatments, either fibrinolysis, thrombectomy, or both.

With the SMOTE resampling technique, the 100% confidence reached for the association of high NIHSS (>20.4) and involvement of the carotid and basilar arteries, with a dire prognosis (exitus), was striking. These results could be extrapolated to previous studies with other assessment methods in which high NIHSS are always associated with a worse prognosis^1)(9)(10. Along with this, it is known that the medial cerebral are statistically the most affected arteries and with the worst NIHSS, with some evolutions being considered malignant strokes.

In summary, in our present study we confirm with machine learning techniques, for the first time, the importance and influence of the association of some traditional predictors, in the stroke prognosis. The analysis of the rules of association of the predictors found in our study will be of particular interest for clinicians, including nurses³⁰, to predict the neurological status of the post-stroke patient.

CONCLUSIONS

Nurses serve in a critical role during a patients’ stroke management and treatment by providing prognosis, communication and care prior to, during and after medical procedures. As a consequence, nurses are often asked, to predict factors that influence post-stroke outcome. In the light of our results and previous publications, nursing staff can successfully improve stroke outcomes, by adequately managing the stroke predictors, obtained not only by means of pure and traditional clinical data, but also by machine learning methods.