Is Long Sleep Associated 
With Increased Risk of Stroke?

SEATTLE—Long sleep duration increases stroke risk, compared with average and short sleep durations, according to an investigation presented at the 29th Annual Meeting of the Associated Professional Sleep Societies. The conclusion results from a Bayesian belief network analysis, which is an uncommon method in neurologic research.

The literature on the association between sleep duration and stroke is equivocal, said Azizi Seixas, PhD, a postdoctoral fellow working with Drs. Girardin Jean-Louis and Gbenga Ogedgebe at the New York University School of Medicine and Dwayne Henclewood (biostatistical support). Dr. Seixas and his colleagues decided to investigate whether short sleep duration (defined as less than seven hours per 24 hours) or long sleep duration (defined as greater than nine hours per 24 hours) is more strongly associated with stroke by comparing logistic regression and Bayesian belief models. The latter method of data analysis allows researchers to interact with their data and update probabilities as new evidence emerges. It also permits researchers to examine bidirectional and omnidirectional relationships between variables.

COMPARING TWO ANALYTICAL METHODS

Dr. Seixas and colleagues conducted a systematic review of behavioral, clinical, individual, familial, and community stroke risk factors using the National Health Interview Survey (NHIS). Data for 288,888 people allowed the researchers to derive a list of 34 risk factors for stroke, including age, sex, hypertension, alcohol consumption, smoking, and kidney disease.

The investigators then analyzed the data using logistic regression and Bayesian belief networks. To conduct the latter analysis, the researchers had to discretize the data (ie, to create multinomial and dichotomous data). Dr. Seixas and colleagues next conducted supervised and unsupervised analyses. In unsupervised learning, a computer performed various evidence-based permutations to find the best-fit model. In supervised learning, the researchers identified stroke as the target variable, and the computer modeled all 34 factors to fit stroke. Finally, the group performed 
a contingent predictive evaluation and an observed adapted 
questionnaire analysis to identify the risk factors most correlated with stroke.

Bayesian Model Was a Better Fit

Approximately 3% of the population had stroke. Logistic regression analysis indicated that long sleep was 
associated with stroke, although the large number of variables decreased the finding’s statistical significance. In an examination of fixed probabilities, the probability of stroke 
was about 3.31% for people with short sleep duration, but long sleep duration increased stroke risk 
by 4.63%.

In contrast to the logistic regression analysis, the Bayesian belief network model did not keep the 34 risk factors constant. This method suggested a J-shaped relationship between long and short sleep durations and indicated that long sleep duration increased stroke risk by a factor of 2.5, while short sleep duration increased stroke risk by a factor of 1.24.

The r² of the logistic regression model was 0.24, and the r² of the Bayesian belief network model was 0.73. This difference “suggests that the Bayesian belief network method is a better model fit,” said Dr. Seixas. “It also has a higher level of accuracy due to this multiple iterative process. Sleep medicine has a significant opportunity to utilize this method as a way of advancing our field.” The logistic regression model could accommodate only 20 of the 34 variables, while the Bayesian belief network model could accommodate all of them.

Among the study’s limitations are that the cross-sectional design of the NHIS may make it difficult to support causal claims about the variables. In addition, the study variables were self-reported, and the data did not include objective biomarkers or objective measures of sleep duration.

The investigation nevertheless suggests that “we could use machine learning, particularly Bayesian statistics, as a way of providing opportunities to assess associations, predictions, and risk-stratification models,” said Dr. Seixas. In future experiments, he and his colleagues intend to identify the pathophysiologic mechanisms that increase long sleepers’ risk of stroke. They also plan to use larger data sets and prospective models in future studies.

This research was supported by the NINDS under Award Number U54NS081765.

—Erik Greb

SEATTLE—Long sleep duration increases stroke risk, compared with average and short sleep durations, according to an investigation presented at the 29th Annual Meeting of the Associated Professional Sleep Societies. The conclusion results from a Bayesian belief network analysis, which is an uncommon method in neurologic research.

The literature on the association between sleep duration and stroke is equivocal, said Azizi Seixas, PhD, a postdoctoral fellow working with Drs. Girardin Jean-Louis and Gbenga Ogedgebe at the New York University School of Medicine and Dwayne Henclewood (biostatistical support). Dr. Seixas and his colleagues decided to investigate whether short sleep duration (defined as less than seven hours per 24 hours) or long sleep duration (defined as greater than nine hours per 24 hours) is more strongly associated with stroke by comparing logistic regression and Bayesian belief models. The latter method of data analysis allows researchers to interact with their data and update probabilities as new evidence emerges. It also permits researchers to examine bidirectional and omnidirectional relationships between variables.

COMPARING TWO ANALYTICAL METHODS

Dr. Seixas and colleagues conducted a systematic review of behavioral, clinical, individual, familial, and community stroke risk factors using the National Health Interview Survey (NHIS). Data for 288,888 people allowed the researchers to derive a list of 34 risk factors for stroke, including age, sex, hypertension, alcohol consumption, smoking, and kidney disease.

The investigators then analyzed the data using logistic regression and Bayesian belief networks. To conduct the latter analysis, the researchers had to discretize the data (ie, to create multinomial and dichotomous data). Dr. Seixas and colleagues next conducted supervised and unsupervised analyses. In unsupervised learning, a computer performed various evidence-based permutations to find the best-fit model. In supervised learning, the researchers identified stroke as the target variable, and the computer modeled all 34 factors to fit stroke. Finally, the group performed 
a contingent predictive evaluation and an observed adapted 
questionnaire analysis to identify the risk factors most correlated with stroke.

Bayesian Model Was a Better Fit

Approximately 3% of the population had stroke. Logistic regression analysis indicated that long sleep was 
associated with stroke, although the large number of variables decreased the finding’s statistical significance. In an examination of fixed probabilities, the probability of stroke 
was about 3.31% for people with short sleep duration, but long sleep duration increased stroke risk 
by 4.63%.

In contrast to the logistic regression analysis, the Bayesian belief network model did not keep the 34 risk factors constant. This method suggested a J-shaped relationship between long and short sleep durations and indicated that long sleep duration increased stroke risk by a factor of 2.5, while short sleep duration increased stroke risk by a factor of 1.24.

The r² of the logistic regression model was 0.24, and the r² of the Bayesian belief network model was 0.73. This difference “suggests that the Bayesian belief network method is a better model fit,” said Dr. Seixas. “It also has a higher level of accuracy due to this multiple iterative process. Sleep medicine has a significant opportunity to utilize this method as a way of advancing our field.” The logistic regression model could accommodate only 20 of the 34 variables, while the Bayesian belief network model could accommodate all of them.

Among the study’s limitations are that the cross-sectional design of the NHIS may make it difficult to support causal claims about the variables. In addition, the study variables were self-reported, and the data did not include objective biomarkers or objective measures of sleep duration.

The investigation nevertheless suggests that “we could use machine learning, particularly Bayesian statistics, as a way of providing opportunities to assess associations, predictions, and risk-stratification models,” said Dr. Seixas. In future experiments, he and his colleagues intend to identify the pathophysiologic mechanisms that increase long sleepers’ risk of stroke. They also plan to use larger data sets and prospective models in future studies.

This research was supported by the NINDS under Award Number U54NS081765.

—Erik Greb

SEATTLE—Long sleep duration increases stroke risk, compared with average and short sleep durations, according to an investigation presented at the 29th Annual Meeting of the Associated Professional Sleep Societies. The conclusion results from a Bayesian belief network analysis, which is an uncommon method in neurologic research.

The literature on the association between sleep duration and stroke is equivocal, said Azizi Seixas, PhD, a postdoctoral fellow working with Drs. Girardin Jean-Louis and Gbenga Ogedgebe at the New York University School of Medicine and Dwayne Henclewood (biostatistical support). Dr. Seixas and his colleagues decided to investigate whether short sleep duration (defined as less than seven hours per 24 hours) or long sleep duration (defined as greater than nine hours per 24 hours) is more strongly associated with stroke by comparing logistic regression and Bayesian belief models. The latter method of data analysis allows researchers to interact with their data and update probabilities as new evidence emerges. It also permits researchers to examine bidirectional and omnidirectional relationships between variables.

COMPARING TWO ANALYTICAL METHODS

Dr. Seixas and colleagues conducted a systematic review of behavioral, clinical, individual, familial, and community stroke risk factors using the National Health Interview Survey (NHIS). Data for 288,888 people allowed the researchers to derive a list of 34 risk factors for stroke, including age, sex, hypertension, alcohol consumption, smoking, and kidney disease.

The investigators then analyzed the data using logistic regression and Bayesian belief networks. To conduct the latter analysis, the researchers had to discretize the data (ie, to create multinomial and dichotomous data). Dr. Seixas and colleagues next conducted supervised and unsupervised analyses. In unsupervised learning, a computer performed various evidence-based permutations to find the best-fit model. In supervised learning, the researchers identified stroke as the target variable, and the computer modeled all 34 factors to fit stroke. Finally, the group performed 
a contingent predictive evaluation and an observed adapted 
questionnaire analysis to identify the risk factors most correlated with stroke.

Bayesian Model Was a Better Fit

Approximately 3% of the population had stroke. Logistic regression analysis indicated that long sleep was 
associated with stroke, although the large number of variables decreased the finding’s statistical significance. In an examination of fixed probabilities, the probability of stroke 
was about 3.31% for people with short sleep duration, but long sleep duration increased stroke risk 
by 4.63%.

In contrast to the logistic regression analysis, the Bayesian belief network model did not keep the 34 risk factors constant. This method suggested a J-shaped relationship between long and short sleep durations and indicated that long sleep duration increased stroke risk by a factor of 2.5, while short sleep duration increased stroke risk by a factor of 1.24.

The r² of the logistic regression model was 0.24, and the r² of the Bayesian belief network model was 0.73. This difference “suggests that the Bayesian belief network method is a better model fit,” said Dr. Seixas. “It also has a higher level of accuracy due to this multiple iterative process. Sleep medicine has a significant opportunity to utilize this method as a way of advancing our field.” The logistic regression model could accommodate only 20 of the 34 variables, while the Bayesian belief network model could accommodate all of them.

Among the study’s limitations are that the cross-sectional design of the NHIS may make it difficult to support causal claims about the variables. In addition, the study variables were self-reported, and the data did not include objective biomarkers or objective measures of sleep duration.

The investigation nevertheless suggests that “we could use machine learning, particularly Bayesian statistics, as a way of providing opportunities to assess associations, predictions, and risk-stratification models,” said Dr. Seixas. In future experiments, he and his colleagues intend to identify the pathophysiologic mechanisms that increase long sleepers’ risk of stroke. They also plan to use larger data sets and prospective models in future studies.

This research was supported by the NINDS under Award Number U54NS081765.

—Erik Greb