Abstract
Background
Post-stroke heart rate (HR) and heart rate variability (HRV) changes have been proposed as outcome predictors after stroke. We used data lake-enabled continuous electrocardiograms to assess post-stroke HR and HRV, and to determine the utility of HR and HRV to improve machine learning-based predictions of stroke outcome.
Methods
In this observational cohort study, we included stroke patients admitted to two stroke units in Berlin, Germany, between October 2020 and December 2021 with final diagnosis of acute ischemic stroke or acute intracranial hemorrhage and collected continuous ECG data through data warehousing. We created circadian profiles of several continuously recorded ECG parameters including HR and HRV parameters. The pre-defined primary outcome was short-term unfavorable functional outcome after stroke indicated through modified Rankin Scale (mRS) score of > 2.
Results
We included 625 stroke patients, 287 stroke patients remained after matching for age and National Institute of Health Stroke Scale (NIHSS; mean age 74.5 years, 45.6% female, 88.9% ischemic, median NIHSS 5). Both higher HR and nocturnal non-dipping of HR were associated with unfavorable functional outcome (p < 0.01). The examined HRV parameters were not associated with the outcome of interest. Nocturnal non-dipping of HR ranked highly in feature importance of various machine learning models.
Conclusions
Our data suggest that a lack of circadian HR modulation, specifically nocturnal non-dipping, is associated with short-term unfavorable functional outcome after stroke, and that including HR into machine learning-based prediction models may lead to improved stroke outcome prediction.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
Introduction
Stroke is a significant public health burden, leading to considerable disability and mortality worldwide [1]. Generally, 45.7% of stroke patients suffer from unfavorable long-term outcome after 3 years (modified Rankin Scale (mRS) > 2) and 5.4% die in-hospital [2]. The main culprits are non-modifiable factors such as stroke severity and pre-stroke disability, but also in-hospital complications like pneumonia, cardiac changes and increased intracranial pressure [2, 3]. Therefore, accurate prediction of stroke outcome upon admission is essential for identifying patients at risk for an unfavorable course and implementing appropriate interventions.
Cardiac metrics like heart rate (HR) and heart rate variability (HVR) have been described in several studies as a proposed clinical parameter for outcome prediction [4,5,6,7]. Changes of HR and HRV in stroke reflect dysregulation of the autonomic nervous system (ANS), which includes decreased parasympathetic activity as well as increased sympathetic activity. This dysregulation can even lead to myocardial injury in the absence of coronary artery disease (CAD) [3, 8, 9]. In stroke, HR and HRV changes can manifest specifically in their circadian rhythm [10]. Studies have shown that higher HR upon admission correlates with poor stroke outcome [11,12,13,14,15,16] and that reduced HRV is associated with higher mortality after stroke [17], suggesting that these metrics could be useful additional outcome predictors. However, current evidence is lacking on whether the addition of HR variables improves stroke outcome prediction models [12]. This shortcoming may have several reasons. First, the time points of outcome assessment have been highly variable in previous studies, ranging from 24 h to 7 years after stroke, which complicates a direct comparison. Second, the precise calculation method of HRV measures differs between studies and is susceptible to noise and artifacts, as well as pharmacological interventions (e. g. antiarrhythmics) [8, 18]. Third, most studies have relied on only a single, brief HR/HRV measurement, usually upon admission, which neglects the well-known circadian modulations of these metrics. There is evidence that the information contained in the entire circadian HR pattern might add additional information beyond single HR or short-term HRV measurements: in hypertensive individuals, a blunted HR dip during sleep was significantly associated with all-cause mortality after up to 14 years of follow-up showing a strong linear correlation independently from other variables [19]. At this stage, it is unclear whether this applies to stroke patients and assessment of circadian HRV changes is not routine.
With the advent of data lake capabilities, it is now possible to store and analyze comprehensive data from stroke patients, including continuous electrocardiography (ECG) over multiple days. The use of these monitoring data and their circadian dynamics can provide a deeper understanding of cardiac and autonomic nervous system (ANS) function in stroke and its utility to improve outcome prediction accuracy. Automated analysis of multi-day in-hospital data to detect atrial fibrillation (AF) and other arrhythmias has already found widespread use. Likewise, this study aims to explore the possibility of using the same approach to monitor HR and HRV over extended periods of time to enhance our understanding of cardiac and ANS function in stroke and improve outcome prediction accuracy to overcome the limitations of single time point measurements.
In this study, we aimed to investigate (1) whether circadian changes in HR and HRV in the first 48 h after stroke are associated with short-term stroke outcome using routinely collected ECG data accessible from a comprehensive data lake structure, and (2) whether existing machine learning (ML) models for stroke outcome prediction could be improved by adding HR as a variable.
Materials and methods
Dataset
All patients diagnosed with nontraumatic intracerebral hemorrhage (ICD-10: I61.-) or ischemic stroke (ICD-10: I63.-) admitted to one of two separate stroke units at Charité Universitätsmedizin Berlin, Berlin, Germany, in the period between October 2020 and December 2021 were initially selected. The two stroke units totalling 20 monitoring beds allowed data transfer and integration into the Data Warehouse Connect (DWC) system (Philips) for long-term storage of monitoring data. The Charité/BIH (Berlin Institute of Health) Health Data Lake (HDL), a Hadoop-based platform that allows storage of a multitude of clinical, epidemiological, laboratory, and monitoring data was used for further integration and analysis. Usage and analysis of the data were approved by the Institutional Review Board of Charité Universitätsmedizin Berlin. The ECG data were recorded with Philips MP30 and MP50 monitors and stored in the data lake. We collected all beat-to-beat intervals from heart beats marked as normal (i.e. sinus rhythm) for up to 48 h after admission. To ensure a meaningful comparison of the HRV measures between subjects, we excluded patients with implanted pacemakers (by excluding beats marked as paced), as well as patients with a known AF diagnosis. In both cases, HRV can be skewed significantly. Other cardiac diseases were not considered for patient exclusion. Beyond ECG measures, the data lake also included a comprehensive set of additional parameters from each patient, including laboratory values, clinical scores, and diagnoses.
Calculation of heart rate and heart rate variability metrics
To calculate the time series-based metrics, the signal was first divided into 5-min segments and a median value was determined for each segment. For every patient, all available signal within the first 48 h post-admission was evaluated. Therefore, assuming the full 48 h of data were available, a patient will have 12 5-min segments per hour and day, resulting in a maximum of 24 5-min segments for each individual hour of day over the 2-day period post-admission. For a more robust estimate, hourly values were only considered if at least 3 of the possible 24 5-min segments were available for the respective hour. HR was calculated as the number of beats per minute. Five HRV measures were considered in this study: standard deviation of beat-to-beat intervals (SDNN), root mean square of successive differences (RMSSD), low frequency power (LF, 0.04–0.15 Hz), high frequency power (HF, 0.15–0.4 Hz) and LF/HF. To calculate frequency domain metrics (LF, HF), the data were resampled by a factor of 4 and interpolated using the UnivariateSpline function of Python’s scipy library. Subsequently, the power spectral density (PSD) was estimated by applying a periodogram function for LF and HF areas, respectively.
Risk factors and predictive features used in machine learning and patient matching
Following previous work [2], an ensemble of risk factors was assessed to provide a benchmark for the ML-powered outcome prediction. Epidemiological factors included age (numeric) and sex (binary value). Other clinical parameters and scores included National Institute of Stroke Scale (NIHSS) at admission (numeric) which quantifies stroke severity, hypertension, diabetes, previous myocardial infarction, known CAD and stroke-associated pneumonia (all binary values). Following previous work [20], serial measurements of high-sensitivity cardiac troponin (assay characteristics: high-sensitivity troponin T, Roche Elecsys, Gen 5; 99th percentile upper reference limit = 14 ng/l; 10% coefficients of variation (CV) precision = 13 ng/l; limit of detection = 5 ng/l) were categorized into three groups: 0—no elevation (below 14 ng/l); 1—at least one elevated sample (above 14 ng/l); 2—acute elevation (one sample above 14 ng/l and a second sample with a 20% increase or decrease in respect to the first sample). Additionally, the glomerular filtration rate (GFR, numeric) was included. Bearing in mind previous work on functional dependence and mRS nomenclature [21], we dichotomized patient outcome into two groups. The first group consisted of patients considered able to live independently, defined by mRS 0–2 upon discharge. The second group included all patients with varying degrees of dependency (including death), i.e., mRS 3–6 upon discharge. From here on out, the two groups are referred to as patients with good vs. unfavorable outcome, respectively. As expected, the two patient groups displayed a high imbalance in age and NIHSS distributions. Given that HR and HRV are known to be age- and NIHSS-dependent, we matched patients according to age and NIHSS to avoid bias [22]. Patients were matched with a propensity score matching (PSM) algorithm (Python library PsmPy) using the k-nearest neighbors method without match replacement and a caliper of 0.2.
Machine learning
We analyzed two ML models with varying degrees of complexity to assess the relevance of HR measures to improve stroke outcome prediction and to validate the robustness of our approach. Specifically, we employed supervised ML to predict the outcome in two classes, good and unfavorable outcome, based on mRS values at discharge. We assessed algorithm performance using a nested cross-validation approach where the dataset was randomly split 4:1 into training and testing sets in a way that the folds preserve the percentage of samples for each class. In a nested fivefold cross-validation, the model’s hyperparameters were tuned and an average model performance calculated over the five folds. This nested cross-validation allowed us to estimate an unbiased performance of the model’s ability to generalize on held-out test sets that were not part of the hyperparameter tuning. This process was repeated 50 times in random shuffles. The hyperparameter ranges used for tuning are displayed in Suppl. Table 1. Data features were normalized by removing the mean and scaling to unit variance. The models included Logistic Regression and CatBoost. CatBoost is a non-linear model based on gradient boosting, which attempts to solve for categorical features using a permutation-driven algorithm. Compared to the previous models, gradient boosting is a more advanced framework for solving regression and classification tasks. For classification, the prediction model is built as an ensemble of weak classification models (classifiers), typically in the form of decision trees. All weak classifiers are then integrated into one cumulative model, that performs better than any of its individual classifiers. In this study, gradient boosting was implemented using the CatBoost library developed by Yandex in 2017.
Machine learning interpretability
In medical applications, the interpretability of ML frameworks is crucial to improve the understanding and acceptance of ML results. In linear models, the models’ coefficients are commonly interpreted as the global feature importance of the respective variables. As such, they may refer to the usefulness of each feature for the prediction. This holds true, if the variables are not exhibiting collinearity, as is the case in this study. For non-linear models a different approach is required to extract a similar relative metric. One recent strategy was derived from game theory and is known as Shapley Additive Explanations. The idea is to compare how well a model performs with and without a feature for every combination of features, therefore calculating a local feature importance for every permutation, as opposed to the global importance that is derived from coefficients and constitutes a single ranking of all features for the model (Table 1). Nevertheless, a global measure can be gained by accumulating local features for every data point. For comparability across different models, both feature importance metrics were normalized (i.e., scaled to unit norm).
Performance assessment and statistical evaluation
ML model performance was assessed on the held-out test sets by calculating the area under the receiver-operator-characteristics (AUC). We reported the mean over the number of shuffles. For HR and HRV measures, statistical independence between groups for each hour of the day was tested by Mann–Whitney-U test with additional Bonferroni correction for multiple comparison (24 individual hours). A two-sample t test was used to compare ML models with and without HR metrics.
Results
We identified 1046 eligible patients across two Charité stroke units admitted between October 2020 and December 2021 matching our diagnosis criteria (Fig. 1). From the initial selection, 421 patients were excluded due to insufficient numbers of normal heart beats labeled for accurate determination of HR/HRV metrics (n = 41), missing data points such as NIHSS, mRS or laboratory parameters like GFR (n = 101) and patients who were diagnosed with AF either during the current hospital stay or in the past (n = 279). We explicitly excluded patients with AF due to the potential impact on HRV metrics, but later checked consistency of our main findings for all patients, including patients with AF. Because HRV metrics and ML performance are known to be highly dependent on age and NIHSS, the remaining 625 patients were matched according to age and NIHSS between groups with good vs. unfavorable outcome, reducing the number of patients considered in the subsequent analysis to 287 (255 with ischemic stroke, 32 with intracranial hemorrhage; Fig. 1). The main patient characteristics are displayed in Table 2.
Circadian dynamics in HR and HRV of patients with good (blue) vs. unfavorable health state (red) after stroke. HR values are significantly lower in the good outcome group with a pronounced dip during the night (gray area). Values indicate means with standard deviation. *Indicates p < 0.05 for difference between respective individual hours, Mann–Whitney-U-Test, Bonferroni corrected
Circadian patterns of HR, HRV and their association with functional stroke outcome
Figure 2 shows HR and HRV metrics for patients with good (blue) and unfavorable outcome (red) after stroke. HR exhibited a significant difference between the two groups with good outcome generally being associated with a lower HR for several hours of the circadian time course. This difference was most pronounced during night hours. We quantified our findings by comparing both groups average HR decline during night hours (22:00–05:00) to a daytime average (10:00–17:00), yielding an average decline of 10.5 ± 2.1 % for patients with good outcome and an average decline of 6.3 ± 1.8 % for patients with an unfavorable outcome. Where the error is the standard error of the mean. Apart from 1 h in the LF/HF time course, no hourly differences were observed for HRV measures.
Nocturnal non-dipping of HR differentiated good from unfavorable stroke outcome at discharge robustly, also when patients were only matched for age, when patients with atrial fibrillation (AF) were included, and when patients were not matched for either age or NIHSS (Fig. 3).
Nocturnal non-dipping of HR differentiates good from unfavorable stroke outcome at discharge. Nocturnal non-dipping of HR as a distinguishing feature of unfavorable outcome is robustly observed when patients are only matched for age (left), when patients with atrial fibrillation (AF) are included (middle) and when patients are not matched for either age or NIHSS (right). PSM: propensity score matching. Data displayed as mean with standard deviation. *Indicates p < 0.05 for difference between respective individual hours, Mann–Whitney-U-Test, Bonferroni corrected
The robust observation of “dippers” vs. “non-dippers” may potentially serve to further categorizes these different patient groups. To provide a potential tool to differentiate the two groups, we quantified the HR dipping for the two different groups by comparing the average HR during nighttime (22:00–05:00) to the HR average during daytime (10:00–17:00). Patients with an unfavorable outcome exhibited a less pronounced dip in HR of 6.3 ± 1.8% on average. In contrast, patients with a good outcome exhibited a more pronounced dip in HR of 10.5 ± 2.1% on average when comparing day- to nighttime (10:00–17:00 vs. 22:00–05:00).
Prediction of stroke outcome
Next, we investigated whether the difference in HR between groups with good vs. unfavorable outcome would improve ML-powered outcome prediction. To this end, we added nocturnal HR, i.e., the mean HR value from 10 pm to 5 am (Fig. 2, gray area), as an independent parameter to the listed clinical, epidemiological and laboratory factors in predicting stroke outcome (Table 2). After adding HR, the prediction performance increased for both models when applied to non-matched data (AUC 0.851 ± 0.005 logistic regression, AUC 0.854 ± 0.005 CatBoost) and when applied to age- and NIHSS-matched data (AUC 0.692 ± 0.012 logistic regression, AUC 0.660 ± 0.013 CatBoost; Fig. 4). Analyzing individual feature importance consistently indicated a high relevance of the HR parameter in all models (Fig. 4, bottom). In conclusion, HR improved the overall performance of ML and was considered a useful additional feature in all models.
Nocturnal HR non-dipping improves stroke outcome prediction as an independent parameter. Top: results for non-matched patients. Bottom: results of patients matched by age and NIHSS. Comparison of prediction performance between feature groups with (orange) and without (blue) utilization of HR (beats per minute, bpm) for prediction and corresponding feature importance for the ML models. For every model type, HR feature importance ranks highly, implying an effective inclusion into the prediction decision. AUC area under the receiver-operator-characteristics, PSM propensity score matching. Data displayed as mean with standard deviation. *Indicates p < 0.05; **p < 0.001; standard two-sample location t test for 50 shuffles of nested cross-validation
Finally, we performed a sub-analysis excluding patients diagnosed with intracranial hemorrhage. All results remained robust (Supplemental Table S1, Figures S1 and S2).
Discussion
By analyzing continuous, data lake-enabled ECGs, we investigated whether good vs. unfavorable outcome after stroke was associated with different circadian HR/HRV profiles and whether these metrics could improve outcome prediction through ML. We found that patients with unfavorable functional short-term outcome showed an overall higher HR with a distinct lack of nocturnal HR-dipping, whereas HRV measures did not differ between the groups. We also found that incorporating HR into the ML models significantly improved the performance of the short-term outcome prediction models compared with models using conventional prognostic factors including stroke severity, age, sex, cardiac history [2, 23]. To the best of our knowledge, no other research to date has used continuous, circadian monitoring data of HR patterns for ML-based outcome prediction.
Our observation that higher HR is associated with unfavorable outcome is in line with several previous studies: resting HR in stroke patients is associated with overall mortality, acceleration of post-stroke cognitive decline and cardiovascular morbidity, especially in men, but not stroke recurrence [11, 13, 15]. Although comparability between these studies is limited, HR is still recognized as an independent risk factor for cardiovascular morbidity and mortality even after adjusting for other risk factors like dyslipidemia, and hypertension [15]. This is in line with our finding that including HR into ML models of stroke outcome prediction significantly improved their performance.
Numerous trials showed an independent relationship between HR and stroke, even after adjusting for the most common cardiovascular risk factors [24,25,26]. This relationship is likely due to the overactivity of the sympathetic nervous system, which seems to predominantly affect the first three days after stroke [15, 27, 28]. Additionally, approximately 50% of stroke patients suffer from insomnia [29]. It is hypothesized that ANS dysfunction is exacerbated during sleep and that sleep disruption after stroke leads to early clinical deterioration through ANS-mediated arrhythmias and blood pressure fluctuations [3]. Hence, we believe it is valid to view HR and its circadian modulation as a surrogate for post-stroke ANS dysfunction. Crucially though, our study also showed a distinct lack of nocturnal HR dipping in patients with unfavorable functional outcome after stroke; none of the studies discussed above differentiated between diurnal and nocturnal HR. It is unclear whether the associations reported by these studies are due to elevated HR in general or specifically due to the blunted HR dip. It should be mentioned that HR fluctuations could also be influenced by other conditions like post-stroke delirium: however, the retrospective analysis of delirium prevalence and its impact on outcome was not feasible within our cohort due to missing data.
Night-time HR dipping gained more interest around the turn of the millennium, when it was discovered that the physiological night-time dip is reduced in patients with CAD or diabetes [30]. A sizable study of 3957 patients showed a remarkably sturdy linear relationship between blunted HR dip and all-cause mortality in hypertensive patients [19]. Interestingly, awake HR was not associated with mortality, and night-time HR showed only a weak association, indicating that it might indeed be the degree of dipping that matters most. HR non-dipping has never been implicated in stroke outcome prediction, although it was suggested to increase the risk of cardiovascular events including stroke in patients treated for hypertension [31].
The impact of stroke on the ANS has been explored for almost 70 years now [32]: for example, HRV is severely decreased in acute ischemic stroke [33,34,35], likely due to autonomic dysregulation after stroke [36]. Some studies found a depression of parasympathetic activity in the acute and chronic stages of stroke, indicated by lower LF [10, 37], while others found that increased sympathetic activity could be the culprit of post-stroke autonomic dysregulation [38]. This dysregulation can lead to immunosuppression and has even been linked to post-stroke infections like pneumonia [39,40,41]. Other signs of post-stroke ANS imbalance include ECG alterations (QTc prolongation, AF, supraventricular and ventricular tachycardia) [42, 43], but also sleep disruption, altered baroreceptor reflex sensitivity, and reduced left ventricular ejection fraction [28, 44, 45].
HRV has been proposed as an independent marker to predict mortality in stroke patients [4, 46, 47]. However, an observational study investigating the predictive power of HRV regarding stroke outcome (n = 308) showed only an association between SDNN and mortality after 3 months; however, similarly to our study, no association of HRV and mRS after 3 months or 1 year could be identified [17].
We used continuous monitoring data to analyze HRV through a ML model. This might help explain, why our results differ from the bulk of the available literature: digesting large amounts of monitoring data will inherently yield different results from traditionally collected data. It is still common practice to calculate HRV measures using a 5–10 min ECG segment after manually eliminating artifacts and premature heart beats [48]. In our study, however, we harnessed the power of ML to analyze a much longer ECG segment.
There is evidence that post-stroke autonomic dysfunction and cardiac changes might be more pronounced in hemorrhagic than ischemic stroke, possibly due to higher stroke severity as well as time-course and extent of brain injury [3]. Therefore, we have accounted for differences in stroke severity by matching for baseline NIHSS scores. We also performed a sub-analysis excluding patients with intracerebral hemorrhage (ICD I61). The results did not change meaningfully.
Limitations
Several limitations must be noticed in our study. First, the measurement of HRV in an acute clinical setting is challenging: HRV is influenced by numerous factors including intake of cardioactive drugs, time of day, and physical activity levels [49]. Second, the optimal method of HRV measurement is still a topic of debate. Initially, linear statistical methods were heavily employed for its calculation, but the last decade has seen a steady rise in non-linear approaches using methods from chaos theory and fractal analysis. Here, we analyzed a comprehensive, well-established set of HRV metrics including spectral measures. Third, our study covered only a relatively short time span between stroke and evaluation of outcome. On average, mRS was calculated 6–7 days after stroke, which might only partially reflect long-term outcome after stroke. Fourth, we did not include imaging data into our analysis, which could lend valuable insight into the effect of the lesion site on HR and HRV changes. Fifth, some studies reported an association of blood pressure and HR non-dipping with small vessel disease (SVD) [50]. Higher prevalence of SVD in HR non-dippers may help explain the worse outcome we observed in our cohort. Sixth, we have no data on our patients’ HR dipping-status before stroke. We argued that a blunted HR dip may represent post-stroke ANS dysfunction, but it may well be already present before stroke, possibly indicating higher pre-stroke morbidity. This could also help explain the worse outcome these patients experience. Seventh, even after matching patients for NIHSS and age, the differences between groups remained substantial. However, our cohort was too small to run numerous sub-analyses or adjust for more variables, which limits the generalizability of the results. Eighth, we opted to use outcome data at discharge, because due to the fairly recent establishment of the utilized data warehouse, long-term outcome data is not yet available. Finally, the reported study was a single-center (though multi-site) study, validation of the presented data in other stroke care centers will be necessary.
Conclusion
Our study showed a distinct pattern of non-dipping HR in patients with short-term unfavorable functional after stroke. This might warrant closer monitoring of HR non-dippers on stroke units and possibly the use of negative chronotropes such as beta-blockers. But whether lowering HR pharmacologically ameliorates stroke outcome is speculative at best and should be thoroughly investigated in further research. HRV measurement in acute stroke patients has proven difficult due to its highly unstable nature that is influenced by many immutable factors in the acute clinical setting. However, a plethora of research has been conducted to simplify its measurement, so there is valid hope that this could lead to increased deployment of this valuable ANS metric in clinical medicine. One could also imagine the inclusion of an automated HR dipping analysis through ECG monitoring on stroke units, alerting clinicians about post-stroke ANS dysfunction.
In conclusion, the availability of large-scale data through data warehousing now allows for a more comprehensive investigation of circadian and other long-term dynamics in cardiac activity and their relationship to stroke outcome. We here report an association between nocturnal non-dipping of HR and unfavorable outcome at discharge in patients with acute stroke. The addition of HR to prediction models based on ML resulted in a significant improvement of prediction performance.
Data availability
The datasets generated and analyzed during the current study are not publicly available due to privacy or ethical restrictions but may be available from the corresponding author on reasonable request. Access to the data will be granted to qualified researchers for purposes of reproducing the results or replicating the procedure.
References
Feigin VL, Stark BA, Johnson CO et al (2021) Global, regional, and national burden of stroke and its risk factors, 1990–2019: a systematic analysis for the Global Burden of Disease Study 2019. Lancet Neurol 20(10):795–820. https://doi.org/10.1016/S1474-4422(21)00252-0
Koennecke HC, Belz W, Berfelde D et al (2011) Factors influencing in-hospital mortality and morbidity in patients treated on a stroke unit. Neurology 77(10):965. https://doi.org/10.1212/WNL.0b013e31822dc795
Scheitz JF, Nolte CH, Doehner W, Hachinski V, Endres M (2018) Stroke-heart syndrome: clinical presentation and underlying mechanisms. Lancet Neurol 17(12):1109–1120. https://doi.org/10.1016/S1474-4422(18)30336-3
Bassi A, Colivicchi F, Santini M, Caltagirone C (2007) Cardiac autonomic dysfunction and functional outcome after ischaemic stroke. Eur J Neurol 14(8):917–922. https://doi.org/10.1111/j.1468-1331.2007.01875.x
Brott T, Adams HP Jr, Olinger CP et al (1989) Measurements of acute cerebral infarction: a clinical examination scale. Stroke 20(7):864–870. https://doi.org/10.1161/01.str.20.7.864
Graff B, Gąsecki D, Rojek A et al (2013) Heart rate variability and functional outcome in ischemic stroke: a multiparameter approach. J Hypertens 31(8):1629–1636. https://doi.org/10.1097/HJH.0b013e328361e48b
Mäkikallio AM, Mäkikallio TH, Korpelainen JT, Sotaniemi KA, Huikuri HV, Myllylä VV (2004) Heart rate dynamics predict poststroke mortality. Neurology 62(10):1822–1826. https://doi.org/10.1212/01.wnl.0000125190.10967.d5
Jimenez-Ruiz A, Racosta JM, Kimpinski K, Hilz MJ, Sposato LA (2021) Cardiovascular autonomic dysfunction after stroke. Neurol Sci 42(5):1751–1758. https://doi.org/10.1007/s10072-021-05128-y
Korpelainen JT, Sotaniemi KA, Huikuri HV, Myllyä VV (1996) Abnormal heart rate variability as a manifestation of autonomic dysfunction in hemispheric brain infarction. Stroke 27(11):2059–2063. https://doi.org/10.1161/01.str.27.11.2059
Tobaldini E, Proserpio P, Oppo V et al (2020) Cardiac autonomic dynamics during sleep are lost in patients with TIA and stroke. J Sleep Res 29(3):e12878. https://doi.org/10.1111/jsr.12878
Böhm M, Cotton D, Foster L et al (2012) Impact of resting heart rate on mortality, disability and cognitive decline in patients after ischaemic stroke. Eur Heart J 33(22):2804–2812. https://doi.org/10.1093/eurheartj/ehs250
Erdur H, Scheitz JF, Grittner U, Laufs U, Endres M, Nolte CH (2014) Heart rate on admission independently predicts in-hospital mortality in acute ischemic stroke patients. Int J Cardiol 176(1):206–210. https://doi.org/10.1016/j.ijcard.2014.07.001
Lee KJ, Kim BJ, Han MK et al (2020) Effect of heart rate on stroke recurrence and mortality in acute ischemic stroke with atrial fibrillation. Stroke 51(1):162–169. https://doi.org/10.1161/strokeaha.119.026847
Nolte CH, Erdur H, Grittner U et al (2016) Impact of heart rate on admission on mortality and morbidity in acute ischaemic stroke patients - results from VISTA. Eur J Neurol 23(12):1750–1756. https://doi.org/10.1111/ene.13115
Tadic M, Cuspidi C, Grassi G (2018) Heart rate as a predictor of cardiovascular risk. Eur J Clin Invest. https://doi.org/10.1111/eci.12892
Tomii Y, Toyoda K, Suzuki R et al (2011) Effects of 24-hour blood pressure and heart rate recorded with ambulatory blood pressure monitoring on recovery from acute ischemic stroke. Stroke 42(12):3511–3517. https://doi.org/10.1161/strokeaha.111.628586
von Rennenberg R, Krause T, Herm J et al (2021) Heart rate variability and recurrent stroke and myocardial infarction in patients with acute mild to moderate stroke. Original Research. Front Neurol 12:772674. https://doi.org/10.3389/fneur.2021.772674
Halimeh M, Yang Y, Sheehan T et al (2022) Wearable device assessments of antiseizure medication effects on diurnal patterns of electrodermal activity, heart rate, and heart rate variability. Epilepsy Behav 129:108635. https://doi.org/10.1016/j.yebeh.2022.108635
Ben-Dov IZ, Kark JD, Ben-Ishay D, Mekler J, Ben-Arie L, Bursztyn M (2007) Blunted heart rate dip during sleep and all-cause mortality. Arch Intern Med 167(19):2116–2121. https://doi.org/10.1001/archinte.167.19.2116
Stengl H, Ganeshan R, Hellwig S et al (2021) Cardiomyocyte injury following acute ischemic stroke: protocol for a prospective observational cohort study. JMIR Res Protoc 10(2):e24186. https://doi.org/10.2196/24186
Saver JL, Chaisinanunkul N, Campbell BCV et al (2021) Standardized nomenclature for modified Rankin Scale global disability outcomes: consensus recommendations from stroke therapy academic industry roundtable XI. Stroke 52(9):3054–3062. https://doi.org/10.1161/strokeaha.121.034480
Hilz MJ, Moeller S, Akhundova A et al (2011) High NIHSS values predict impairment of cardiovascular autonomic control. Stroke 42(6):1528–1533. https://doi.org/10.1161/strokeaha.110.607721
Zihni E, Madai VI, Livne M et al (2020) Opening the black box of artificial intelligence for clinical decision support: a study predicting stroke outcome. PLoS ONE 15(4):e0231166. https://doi.org/10.1371/journal.pone.0231166
Aune D, Sen A, Óhartaigh B et al (2017) Resting heart rate and the risk of cardiovascular disease, total cancer, and all-cause mortality—a systematic review and dose-response meta-analysis of prospective studies. Nutr Metab Cardiovasc Dis 27(6):504–517. https://doi.org/10.1016/j.numecd.2017.04.004
Fox K, Bousser MG, Amarenco P et al (2013) Heart rate is a prognostic risk factor for myocardial infarction: a post hoc analysis in the PERFORM (Prevention of cerebrovascular and cardiovascular Events of ischemic origin with teRutroban in patients with a history oF ischemic strOke or tRansient ischeMic attack) study population. Int J Cardiol 168(4):3500–3505. https://doi.org/10.1016/j.ijcard.2013.04.206
Zhang D, Wang W, Li F (2016) Association between resting heart rate and coronary artery disease, stroke, sudden death and noncardiovascular diseases: a meta-analysis. CMAJ 188(15):E384-e392. https://doi.org/10.1503/cmaj.160050
Sykora M, Diedler J, Poli S, Rupp A, Turcani P, Steiner T (2010) Blood pressure course in acute stroke relates to baroreflex dysfunction. Cerebrovasc Dis 30(2):172–179. https://doi.org/10.1159/000317105
Yperzeele L, van Hooff RJ, Nagels G, De Smedt A, De Keyser J, Brouns R (2015) Heart rate variability and baroreceptor sensitivity in acute stroke: a systematic review. Int J Stroke 10(6):796–800. https://doi.org/10.1111/ijs.12573
Palomäki H, Berg A, Meririnne E et al (2003) Complaints of poststroke insomnia and its treatment with mianserin. Cerebrovasc Dis 15(1–2):56–62. https://doi.org/10.1159/000067127
Burger AJ, Charlamb M, Sherman HB (1999) Circadian patterns of heart rate variability in normals, chronic stable angina and diabetes mellitus. Int J Cardiol 71(1):41–48. https://doi.org/10.1016/s0167-5273(99)00110-2
Eguchi K, Hoshide S, Ishikawa J et al (2009) Nocturnal nondipping of heart rate predicts cardiovascular events in hypertensive patients. J Hypertens 27(11):2265–2270. https://doi.org/10.1097/HJH.0b013e328330a938
Burch GE, Meyers R, Abildskov JA (1954) A new electrocardiographic pattern observed in cerebrovascular accidents. Circulation 9(5):719–723. https://doi.org/10.1161/01.cir.9.5.719
Barron SA, Rogovski Z, Hemli J (1994) Autonomic consequences of cerebral hemisphere infarction. Stroke 25(1):113–116. https://doi.org/10.1161/01.str.25.1.113
Naver HK, Blomstrand C, Wallin BG (1996) Reduced heart rate variability after right-sided stroke. Stroke 27(2):247–251. https://doi.org/10.1161/01.str.27.2.247
Tokgözoglu SL, Batur MK, Topçuoglu MA, Saribas O, Kes S, Oto A (1999) Effects of stroke localization on cardiac autonomic balance and sudden death. Stroke 30(7):1307–1311. https://doi.org/10.1161/01.str.30.7.1307
Oka H (2017) Heart rate variability and neurological disorders. In: Iwase S, Hayano J, Orimo S (eds) Clinical assessment of the autonomic nervous system. Springer Japan, Tokyo, pp 179–197
Xiong L, Leung HH, Chen XY et al (2013) Comprehensive assessment for autonomic dysfunction in different phases after ischemic stroke. Int J Stroke 8(8):645–651. https://doi.org/10.1111/j.1747-4949.2012.00829.x
Dütsch M, Burger M, Dörfler C, Schwab S, Hilz MJ (2007) Cardiovascular autonomic function in poststroke patients. Neurology 69(24):2249–2255. https://doi.org/10.1212/01.wnl.0000286946.06639.a7
Engel O, Akyüz L, da Costa Goncalves AC et al (2015) Cholinergic pathway suppresses pulmonary innate immunity facilitating pneumonia after stroke. Stroke 46(11):3232–3240. https://doi.org/10.1161/strokeaha.115.008989
Günther A, Salzmann I, Nowack S et al (2012) Heart rate variability—a potential early marker of sub-acute post-stroke infections. Acta Neurol Scand 126(3):189–196. https://doi.org/10.1111/j.1600-0404.2011.01626.x
Prass K, Meisel C, Höflich C et al (2003) Stroke-induced immunodeficiency promotes spontaneous bacterial infections and is mediated by sympathetic activation reversal by poststroke T helper cell type 1-like immunostimulation. J Exp Med 198(5):725–736. https://doi.org/10.1084/jem.20021098
Hjalmarsson C, Bokemark L, Fredriksson S, Antonsson J, Shadman A, Andersson B (2012) Can prolonged QTc and cTNT level predict the acute and long-term prognosis of stroke? Int J Cardiol 155(3):414–417. https://doi.org/10.1016/j.ijcard.2010.10.042
Kallmünzer B, Breuer L, Kahl N et al (2012) Serious cardiac arrhythmias after stroke: incidence, time course, and predictors—a systematic, prospective analysis. Stroke 43(11):2892–2897. https://doi.org/10.1161/strokeaha.112.664318
Ois A, Gomis M, Cuadrado-Godia E et al (2008) Heart failure in acute ischemic stroke. J Neurol 255(3):385–389. https://doi.org/10.1007/s00415-008-0677-1
Palamarchuk I, Kimpinski K, Lippert C, Hachinski V (2013) Nocturnal deterioration after ischemic stroke and autonomic dysfunction: hypothesis and implications. Cerebrovasc Dis 36(5–6):454–461. https://doi.org/10.1159/000356093
Robinson TG, Dawson SL, Eames PJ, Panerai RB, Potter JF (2003) Cardiac baroreceptor sensitivity predicts long-term outcome after acute ischemic stroke. Stroke 34(3):705–712. https://doi.org/10.1161/01.Str.0000058493.94875.9f
Scherbakov N, Barkhudaryan A, Ebner N et al (2020) Early rehabilitation after stroke: relationship between the heart rate variability and functional outcome. ESC Heart Fail 7(5):2983–2991. https://doi.org/10.1002/ehf2.12917
Task Force of the European Society of Cardiology the North American Society of Pacing Electrophysiology (1996) Heart rate variability. Circulation 6:1043–1065
Quintana DS, Heathers JAJ (2014) Considerations in the assessment of heart rate variability in biobehavioral research. Front Psychol 5:805–805. https://doi.org/10.3389/fpsyg.2014.00805
Akhtar N, Al-Jerdi S, Kamran S et al (2021) Night-time non-dipping blood pressure and heart rate: an association with the risk of silent small vessel disease in patients presenting with acute ischemic stroke. Front Neurol 12:719311. https://doi.org/10.3389/fneur.2021.719311
Acknowledgements
The study was supported by NeuroCure Clinical Research Center (NCRC), funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany´s Excellence Strategy EXC 2049-390688087. ME also reports funding from Collaborative Research Center ReTune TRR 295-424778381, BMBF, DZNE, DZHK, EU, Corona Foundation, and Foundation Leducq.
Funding
Open Access funding enabled and organized by Projekt DEAL.
Author information
Authors and Affiliations
Contributions
All authors contributed to the study conception and design. Material preparation, data collection and analysis were performed by AN and CM. The first draft of the manuscript was written by MGK and AN and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.
Corresponding author
Ethics declarations
Conflicts of interest
CM is part of patent applications to detect and predict clinical outcomes and to manage, diagnose, and treat neurological conditions, all of which are outside the submitted work. CHN received research grants from German Ministry of Research and Education, German Center for Neurodegenerative Diseases, German Center for cardiovascular Research, and speaker and/or consultation fees from Abbott, Alexion, Astra Zeneca, Bayer Pharma, Bristol-Myers Squibb, Daiichi Sankyo, Novartis, Pfizer Pharma, Portola and Takeda all outside the submitted work. ME reports grants from Bayer and fees paid to the Charité from Abbott, Amgen, AstraZeneca, Bayer, Boehringer Ingelheim, BMS, Daiichi Sankyo, GSK, Sanofi, Covidien, Novartis, Pfizer, all outside the submitted work. The remaining authors have no conflicts of interest or competing interests to declare.
Ethical approval
Ethics approval for this study was obtained from the Institutional Review Board of the Charité Universitätsmedizin Berlin. The study was conducted in accordance with the ethical principles outlined in the Declaration of Helsinki.
Supplementary Information
Below is the link to the electronic supplementary material.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Nelde, A., Klammer, M.G., Nolte, C.H. et al. Data lake-driven analytics identify nocturnal non-dipping of heart rate as predictor of unfavorable stroke outcome at discharge. J Neurol 270, 3810–3820 (2023). https://doi.org/10.1007/s00415-023-11718-x
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00415-023-11718-x