Over several decades, actigraphy has been used to objectively examine rest and wake periods in a variety of participants. Actigraphs are accelerometers that measure activity (acceleration of motion) levels of the person wearing the device [1-3]. Given its small footprint and lower cost, actigraphy is often used as a simple alternative to traditional, more invasive sleep monitoring methods. Actigraphy is used to measure basic movement activity patterns of individuals noninvasively to assess when they are asleep (rest) or awake (wake) [1-3]. Movement data collected by actigraphs are analyzed using validated algorithms to determine, for each minute of recording, whether the wearer was asleep or awake. At the time of this publication, the vast majority of sleep-wake analysis of movement data collected by actigraphy continues to be done primarily using one of a handful of popular linear regression models. These include Cole-Kripke, University of California San Diego (UCSD), Sadeh, and Philips-Respironics. Though these algorithms are extensively used, they have largely been tested with only small samples of young and healthy individuals [3-11]. There is little research examining the comparison between and validity of these algorithms with a large, demographically diverse sample or with a sample at risk of health conditions. Given the increasing use of actigraphy for a variety of populations, it is critical to evaluate these algorithms across a large, diverse population to give a better representation of their accuracy. Further, it is important to evaluate the accuracy of these algorithms across older adults, as these populations have higher susceptibility to, and high prevalence of health disorders. Given the wide range of studies using actigraphy, it is important to understand actigraphy’s performance when used with these higher-risk populations for sleep-wake assessment [3,12-17]. In addition, given the comorbidity of sleep issues with other health disorders, the use of actigraphy in sleep studies has become more commonplace [3,12-17]. Therefore, it is highly important to assess the accuracy and reliability of traditional actigraphy algorithms with these populations.

There are several significant challenges in collecting a large, diverse dataset. This, in part, is due to the ideal method of validation (for actigraphy), which is polysomnography, the gold standard for sleep assessment. Polysomnography and its analysis are expensive, clinic-based procedures that require extensive equipment and expertise, are highly time-consuming to conduct, and are therefore impractical for measuring sleep across several nights with large samples [8,18,19]. However, an ongoing initiative called the Multi-Ethnic Study of Atherosclerosis (MESA) has conducted a large sleep study (over 2000 participants), including time-synchronized actigraphy and polysomnography data as part of their larger initiative [20,21]. This sleep sample represents a portion of over 6800 participants between the ages of 45 and 84 years, free of cardiovascular disease at baseline, who were monitored longitudinally for subclinical cardiovascular disease [20,21]. This dataset presents a novel opportunity to establish a concrete comparison between conventional algorithms and establish their accuracy across a large population of older adults who are also at risk of health issues. Thus, this paper evaluates traditional actigraphy algorithms using the MESA dataset.

The primary objective of this study was to establish the accuracy of traditional actigraphy algorithms against the gold standard polysomnography using a large established dataset. In doing so, we provide comprehensive foundational benchmarks of the most commonly used traditional actigraphy algorithms. Further, we aimed to also establish performance benchmarks of commonly used actigraphy algorithms across a diverse population with older adults and sleep pathologies. These benchmarks greatly inform research and clinical use of these algorithms. Further, this study provides detailed standards to developers of novel actigraphy technologies and actigraphy analysis methodologies. Based on previous, albeit smaller, sample validation studies, we hypothesized that traditional actigraphy algorithms would detect sleep and wake with high accuracy. In addition, we predicted that additional rescoring of actigraphy would help improve their accuracy over the algorithms alone.

In addition to sleep-wake, researchers and clinicians often examine sleep metrics to evaluate patterns or disruption of sleep patterns. As our secondary goal, we examined each algorithm’s agreement and bias with polysomnography on several commonly studied sleep metrics, including total sleep time (TST), sleep efficiency (SE), and wake after sleep onset (WASO). TST is the total number of minutes an individual sleeps during the night from the first onset of sleep to sleep offset. On the other hand, WASO is the total duration of wake between the first onset of sleep and the sleep offset. Finally, SE is the percentage of time asleep (TST) during the total time spent in bed. As these are key sleep metrics, this study’s results provide detailed information for current research and clinical practice as well as benchmarks for future improvements to current and novel actigraphy methods. We hypothesized that overall, sleep metrics derived from actigraphy algorithm-processed activity would have high agreement with polysomnography [22].

As an exploratory analysis, we looked at both accuracy and sleep metrics of each algorithm and evaluated possible points of failure and variables that impact the performance of actigraphy analyses. Specifically, we examined if accuracy results and sleep metrics would be poorer for subgroups of participants who had sleep disorders. To our knowledge, this is the first study to provide detailed actigraphy performance metrics across large key samples of individuals with sleep problems. These metrics provide nuanced information that will critically facilitate current and future practice. We discuss the implications of these variables, making recommendations for future directions of research.

The dataset sample for this study was derived from the MESA and acquired through the National Sleep Research Resource (NSRR) [20,21]. MESA is a large prospective, community-based study designed to examine the risk factors, prevalence, and progression of cardiovascular disease [20,21]. The MESA study was conducted from 2000 to 2002, with 6814 ethnically diverse men and women aged 45‐84 years who were free of overt cardiovascular disease and recruited from 6 US sites [20,21]. MESA participants underwent periodic core physical assessments at enrollment and 10 years following initial recruitment (2010‐2012) [20,21]. A subset of participants was invited to participate in the MESA ancillary sleep assessment study. This study entailed a home visit (1 night) in which home polysomnography was conducted along with concurrent actigraphy. Polysomnography data were analyzed by one of 3 polysomnologists using standard guidelines, while 2 scorers completed actigraphy analysis by first marking the sleep period and then using automatic scoring software. Scorers completed a MESA-rule–based training and certification before scoring, and postscoring both interscorer reliability was assessed at 2 timepoints for sleep stages (which pertained to the sleep data used for the study). At time point 1 for sleep stage scoring, the interscorer intraclass correlation coefficients (ICC) across n=27 from 9 participants were as follows: Stage 1 was 0.86, Stage 2 was 0.63, Stage 3‐4 was 0.81, and REM was 0.96 [20,21]. At time point 2, interscorer ICC across n=38 from 19 participants were as follows: Stage 1 was 0.74, Stage 2 was 0.81, Stage 3‐4 was 0.79, and REM was 0.93 [20,21]. For this study, we used raw actigraphy activity counts for analysis. For ground truth, we used the matching polysomnography sleep-wake data that was collapsed to wake (coded wake stage) versus all sleep stages collapsed to one “sleep” variable. Only one night of polysomnography data were collected by the original study which was matched and synchronized with the respective actigraphy data (cropped to match).

The data were filtered based on several criteria. The initial dataset contained 2159 samples for actigraphy and 2056 samples for polysomnography. We first excluded samples that did not have concurrent polysomnography data (103 removed). The actigraphy and polysomnography samples were then matched based on a synchronization document provided by NSRR and MESA. Those samples that did not have a corresponding synchronization or were reported as having data issues by NSRR (2 individuals) were excluded, resulting in a sample of 1831 individuals (225 removed). Next, we filtered the dataset based on polysomnography data quality as indicated by the dataset description. We included individuals for whom the quality of polysomnography data was rated as very good or better, as this rating indicated an appropriate number of channels present for accurate sleep-wake evaluation. An additional 14 participants with complete, matched data of fair and good quality were also included. This resulted in a total of 1484 samples (347 removed).

Outlier filtering, or filtering based on actigraphy, was not conducted, as we wished to retain all actigraphy data for a true representation of how actigraphy data are traditionally processed using these algorithms. Given the absence of outlier filtering, we conducted sensitivity analysis to evaluate the effects of these outliers on our overall sample results. We identified outliers in the cropped 30-second epoch synchronized (1-night) actigraphy raw activity data using a 2-step process. The raw 30-second data were used as this was the data used for each algorithm (the binary algorithm results could not be used to identify outliers). Each participant’s individual iIQR, mean, and SD of activity were calculated. For the first level, participants’ IQR was compared to the full sample IQR. Any participants whose range fell outside of 1.5 times below the first quartile and above the first quartile were identified as outliers. For the second step, the z score for each participant was calculated relative to the sample mean and SD. Any participants with a z score of 2 or greater were identified as outliers. A total of 56 participants were identified as outliers using this 2-tier approach. Once these outlier individuals were identified, they were omitted, and Cohen κ, Matthews correlation coefficient (MCC), and confusion matrix metrics were recalculated. The relative change between the full dataset and outlier-omitted results was calculated to be less than 5% across all metrics and algorithms except for the Philips 40 rescored algorithm sensitivity (5.55%) and specificity (7%). However, the relative change was for only the rescored algorithm, close to the 5% threshold, and all other metrics were within the 5% range; therefore, we determined that the outliers did not pose a significant issue. For full comparison metrics, see Multimedia Appendix 1.

Further, polysomnography data were in binary format, and the actigraphy data, once processed through an algorithm, would also be in binary data point format. This would limit the breadth of fluctuation in the main analyses and would allow for direct 1:1 comparison of epochs. In addition to the main analyses, rescoring was also applied, which is the standard method of correcting issues with actigraphy data. In addition, we retained all participants for subsequent analysis of sleep metrics, as we aimed to evaluate the entire spread of sleep metric results for normal sleep, sleep problems, and extreme cases of sleep. As a secondary measure, we added analyses for participants with sleep problems to evaluate these extreme cases separately.

Finally, during data processing, an additional 44 individuals were excluded due to issues synchronizing actigraphy and polysomnography time points, missing data, and issues with processing. The final sample size analyzed was N=1440 (M_age=69.3, SD 9.0 years; n_male=663) individuals (Figure 1). The distribution of included participants is denoted in Table 1.

In addition to the final sample, for exploratory analyses, subsamples for the populations with sleep problems were also selected. Four sleep problem groups, including apnea, individuals who used continuous positive airway pressure (CPAP), insomnia, and restless leg syndrome (RLS), were examined using the same detailed analyses and methods as the full sample (Table 1).

Data processing was conducted using Python 3 with a wide range of libraries and platforms. Before the final analyses, both actigraphy data and polysomnography data were transformed from 30-second epochs to 1 minute to standardize, since many algorithms and previous studies typically present minute-by-minute epochs for actigraphy for easier data analysis and interpretation [22,23]. For the actigraphy data, transformation was done after algorithm results were obtained. For the actigraphy results, once each respective algorithm had processed the data, the actigraphy values from the algorithms were summed to collapse the data to 1-minute epochs. The binarized polysomnography data were also transformed to 1-minute epochs (from 30-second epochs) using the following set of rules. First, if 2 adjacent 30-second epochs were both coded as sleep, the resultant minute would be sleep. Second, if 2 adjacent 30-second epochs were both coded as wake, the resultant minute would be wake. Third, if one of 2 adjacent 30-second epochs were wake, the resultant minute would be wake. Polysomnography transformation to collapse epochs to 1-minute intervals is a common practice for actigraphy comparison [24-27]. With respect to the wake-wins, evaluation of the distribution of pairs was conducted for the polysomnography epochs across all participants. A total of 55.13% of pairs were both asleep, 38.33% were both awake, and only 6.54% were mixed (wake-sleep or sleep-wake). Given the low percentage of mixed pairs, we expected the effect of data transformation to be limited with respect to the wake-wins condition.

All data transformations were again done through Python 3, using a wide range of libraries with coding assistance provided by ChatGPT (OpenAI) to generate Python code for existing Python libraries [28]. ChatGPT was used to generate Python code for file data file preparation and processing (eg, data transformation). Manual data checks in addition to code-based error reporting were conducted to verify the code was working as desired and data were correctly processed. In addition, ChatGPT was used to generate Python code for running statistical analyses. All statistical formulae and analyses were verified within the code. Once code was generated, it was checked by at least one or more of the study authors before use. No machine learning code was generated via ChatGPT, and any actigraphy algorithm code was checked and specified in the Python code exactly as defined by the original algorithm sources. Data were synchronized based on the NSRR and the MESA-provided synchronization guide, which highlighted polysomnography start times relative to actigraphy.

The current study was conducted on a pre-established MESA dataset on a secondary use basis [20,21]. The original MESA study obtained institutional review board ethics approval for each study site of data collection and obtained written consent from all participants [20]. The NSRR resource maintains a detailed approval procedure for dataset acquisition and use [21]. The study dataset was acquired through a data access and use agreement with the NSRR [21]. All data were predeidentified for privacy by NSRR. Ethics approval was acquired from the University of Toronto (REB protocol # 35344) for this study.

Using the large MESA dataset, we established that traditional actigraphy algorithms can, with considerable accuracy, classify sleep and wake. As a first, this study comprehensively provides detailed benchmarks of traditional actigraphy algorithms. This study also provides the first rescoring performance and comparison across a wide range of actigraphy algorithms. This study is among the first to evaluate traditional actigraphy algorithms across a large population of older adults who are also at risk for, or have, health and sleep issues. To our knowledge, this is the first study to provide comprehensive benchmarks and statistical comparisons across an extensive list of the commonly used actigraphy algorithms with the diverse MESA population, while previous studies have focused on novel algorithm development and comparison. In doing so, we provide comprehensive metrics for not only users of these algorithms but also inform future research and practice.

Overall, the ROC-AUC and accuracy results demonstrate that all of the algorithms evaluated in this study are good options for evaluating sleep-wake activity. However, contrary to accuracy, our results for both MCC and Cohen κ demonstrate that traditional algorithms only provide moderate agreement and moderate positive correlations with ground truth polysomnography. These results suggest that though these actigraphy algorithms are a valuable analytic tool, there is a significant area of improvement. With novel technologies as well as improved analytic methods, we can improve our approach to sleep-wake assessment.

Along these lines, our results demonstrate that rescoring of actigraphy algorithms does not substantially improve sleep-wake classification accuracy. Rather, rescoring may significantly diminish the performance (eg, F₁-score) of certain algorithms such as Kripke 2010 and Philips-Respironics. Further, our analyses of sleep problem participants revealed similar findings that rescoring may diminish sleep-wake classification performance. This suggests that researchers and clinicians may safely opt to not rescore and still retain high accuracy.

In addition to the primary results, in-depth ANOVA and post hoc analyses showed significant differences between algorithms despite small differences in accuracy and other metrics. However, examination of effect sizes for MCC, accuracy, and F₁-score demonstrated very small to small effect sizes. Therefore, the significant differences between algorithms could be due to the large sample size of the dataset. The exception was Cohen κ, which demonstrated marginally higher performance for the Kripke 2010 algorithm (small to medium effect sizes). This, along with the Kripke 2010 algorithm demonstrating the highest accuracy, may suggest that this algorithm could be optimal in specific analysis scenarios. However, after rescoring, these effects were reduced with only very small to small effect sizes overall.

Our evaluation of the agreement between actigraphy and polysomnography for sleep metrics TST, SE, and WASO was conducted through Bland-Altman distribution analysis. This analysis demonstrated both over and underestimation of all 3 sleep metrics, underscored by large mean differences between actigraphy and polysomnography. The Bland-Altman distributions also demonstrated considerable systematic bias for SE and WASO. In addition, there were clear patterns that demonstrated agreement decreased as SE decreased or when WASO increased. This was supported by the regression analyses for proportional bias. This suggests that actigraphy may be less precise in accurately measuring sleep metrics in individuals who have irregular sleep or sleep problems. From the traditional algorithms, the Philips algorithm with a threshold of 40 demonstrated the highest agreement with polysomnography for both TST and SE. This may be in part due to the supplied cutoffs for sleep-wake detection, which may allow for the user to adjust the algorithm based on their target population and other factors to improve both sleep-wake scoring and sleep metric estimates [37]. However, an important note is that the MESA data were collected using Philips Actiwatch devices; therefore, it is expected that the Philips algorithms may perform better in these cases than the other algorithms [20,21]. For WASO, the Kripke 2010 algorithm had the highest agreement with polysomnography. This may be due to the larger number of epochs (30) that are assessed by the algorithm to compute a sleep-wake decision. In evaluating more epochs, this algorithm may be more sensitive to shorter waking periods [19]. This may also explain the lower accuracy of this algorithm, as it may be scoring more epochs as wake even when there is true sleep. This suggests a key point for current algorithm use and for future algorithm development, that some variability or ranges of thresholds and longer windows of epoch evaluation may provide better accuracy and improve estimates of sleep metrics.

Another explanation for the bias and heteroscedasticity observed is that the sample evaluated included many individuals who were either at risk of or had health conditions such as the sleep problems highlighted in the sample distributions (Table 1). Though not reported in the NSRR version of the MESA dataset, MESA and publications on this dataset have underscored that this sample includes individuals with hypertension, diabetes, etc [20,48,49]. Therefore, it is expected that these conditions may lead to fragmented and irregular sleep, which may have influenced the actigraph or the algorithm’s ability to accurately measure. Previous studies have noted that individuals with sleep problems or health conditions that may affect sleep demonstrate more variability in actigraphy results of both sleep-wake and sleep metrics [3,12-17]. Therefore, one important consideration was further tests of these algorithms with data from populations with health and sleep issues. Notably, these are the key populations who are recommended or given actigraphs for sleep assessment by researchers and clinicians. Therefore, the precision of their sleep or wake classification and sleep metrics by these algorithms is of great importance.

Along these lines, we first conducted an exploratory analysis of algorithm performance within sleep problem subgroups, specifically examining individuals with insomnia, RLS, apnea, and those who used a CPAP device. We found that traditional actigraphy algorithms do not perform particularly worse; rather, they are just as robust in determining sleep and wake or sleep metrics in individuals with sleep problems compared with the entire sample. These algorithms demonstrated very similar results and patterns to the entire sample with respect to AUC, accuracy, confusion matrix results, MCC, Cohen κ, and agreement (Bland-Altman distributions). However, generally, these algorithms perform well with respect to sleep-wake scoring but not as well for the measurement of specific sleep metrics in contrast to the overall population. Our results show that sleep disorders may skew Bland-Altman distributions as sleep patterns within these disorders deviate from norms. We theorize based on our results that this may result in larger mean differences when compared to ground truth. Based on our results, we theorize that this may create skewness in accuracy in particularly extreme cases. We suggest that clinicians pay attention to severe cases of sleep disorders and take into account our results and the decreased reliability of actigraphy-based sleep metrics such as TST due to outlier sleep patterns.

Overall, these results provide clinicians with nuanced benchmarks for individual algorithm performance on specific sleep disorders. This will allow clinicians to make an informed choice not only based on accuracy, but on sensitivity and specificity metrics to precisely analyze patient sleep data. With mean difference and Bland-Altman plots, clinicians now have benchmarks to provide them with information on how extreme cases may deviate from general trends. Further, both clinicians and manufacturers now have mean difference benchmarks, which reflect the difference between ground truth and actigraphy, allowing for corrections in patient data analyses if needed.

In the future, it would be interesting to compare the precision of traditional actigraphy algorithms across a variety of actigraph and accelerometer devices. Overall, both the main and subgroup results suggest that generally traditional actigraphy algorithms may have poorer agreement with polysomnography in fringe cases. Specifically, cases in which the individual’s sleep may be significantly different from the vast majority of the population. One hypothesis here is that the heteroscedasticity may be correlated to the severity of a sleep problem, which explains the poorer agreement as SE or WASO changes. Further analysis would allow us to gain a better understanding of this phenomenon.

With regard to rescoring and the sleep problem sub-samples, there were mixed results. Traditional use of the Cole-Kripke (1992) [6] algorithm proposed the use of rescoring; first, we investigated rescoring across all included algorithms [40]. Contrary to the belief that rescoring would significantly improve accuracy metrics, rescoring only slightly improved agreement for sleep metrics and may diminish specific performance variables. This could be due to the rescoring of smaller periods of sleep and wake, especially in the aforementioned population with sleep problems. Rescoring may be better suited across multiple nights, which may present more within-participant variability, whereas our sample was only over one night. Rescoring may also smooth out or erroneously change edge cases or cases in which individuals have sleep problems. This would, in part, explain the diminished performance of results. One future direction could be further optimization of rescoring criteria for improvements to sleep-wake detection across algorithms or to better suit individual algorithms specifically. This may also remove the diminished performance observed after rescoring with sleep problem populations. As a current recommendation, we suggest that rescoring be used only when absolutely necessary and that when rescoring, researchers compare results that are both rescored and nonrescored. Our findings suggest that rescoring does not provide any substantial benefits to nonrescored algorithm results. Further, we do not recommend rescoring for sleep disorder population data to preserve accurate sleep-wake scoring.

Overall, these results echo the findings of previous studies which have examined these algorithms. Haghayegh et al [50] examined the performance of the Coke-Kripke, Sadeh, and UCSD algorithms on a sample of 40 healthy adults. Similar to our results, they found accuracies of 85%‐86% across algorithms. Additionally, they also found varying levels of overall bias across algorithms for the algorithms’ estimation of total sleep time, sleep onset latency, and wake after sleep onset. Along these lines, Palotti et al [51] and Jokar et al [52] used a smaller subset of the MESA dataset to compare performance to the Cole-Kripke and Sadeh algorithms to their novel method both for night only and night-day. Cole-Kripke and Sadeh demonstrated mean accuracies ranging from 70%to 85% [53,54].

In addition, this study presented some limitations. Foremost, there has been recent work and novel algorithms for sleep and wake classification. Some notable examples include novel machine learning and other methods to evaluate activity data. These algorithms may be more precise and robust as sleep and wake classification. Though our current study did not directly evaluate these algorithms, previous research has provided some benchmarks for comparison. Palotti et al [51] developed novel Convolutional Neural Network and Long Short-Term Memory machine learning approaches using the MESA dataset for training and testing. Their novel approaches demonstrated better performance (88.2% and 87.7% respectively) than traditional algorithms (Cole-Kripke and Sadeh). In comparison to this study, their novel methods also demonstrate higher performance than traditional algorithms, albeit on a smaller subsample of the MESA dataset. Similarly, Nunes et al [55] developed a domain adversarial convolutional neural network method, trained on the MESA dataset. Their optimal method demonstrated the highest accuracy (80.1%) when compared to other models and traditional algorithms such as Cole-Kripke and Sadeh (71.9% and 69.6%). In contrast to our findings, their novel method could provide stable, more generalizable sleep-wake classification. Again, these results suggest that novel machine learning methods are the future for actigraphy analyses. Given the rapid development of novel actigraphy analytic methods, our future study aim is to examine the performance of newer algorithms and strategies with large testing samples in contrast to our current study findings.

Another limitation of this study is that we only examined sleep problem subgroups as an exploratory aim. Given the wide demographic, there were several other populations within the dataset, such as individuals with diabetes, cardiovascular illness, etc that could explain the lack of agreement or the heteroscedasticity observed. In addition, we did not conduct an analysis examining demographic subgroups such as sex, ethnicity, and age. These groups may have also contributed to the Bland-Altman distributions observed.

Given the similarities of the subgroup results to those of the entire dataset, there may be additional correlation-based analysis, which could help unpack why traditional actigraphy algorithms perform poorly as sleep metrics reach extreme values. As aforementioned, previous research denotes individuals with sleep problems may have lower SE, higher WASO, and a variety of other extreme sleep activity metrics. Therefore, this may be one important limitation of not only the actigraphy algorithms but the activity measurement devices themselves. To address this issue, future research should delve deeper and examine extreme cases and traditional algorithms’ ability to classify sleep and wake and accurately measure sleep metrics. This, in turn, may also allow us to gain a better understanding of specific patterns of sleep and wake in these populations, as well as examine whether specific populations alter the distribution variability observed. In addition, future research should comprehensively assess the performance of traditional and novel algorithms on subsets of the MESA dataset. In doing so, we could identify key markers that may be used to improve analysis performance.

Finally, our study also filtered out fragmented or poor-quality data from the dataset. Even though these data may not be suitable for the current accuracy analysis and comparison, they may highlight important limitations of actigraphy algorithms with respect to the estimation of sleep metrics and the classification of sleep and wake in extreme cases. Therefore, future research should assess a spectrum of poor- and good-quality data to gain a complete understanding of the performance of traditional actigraphy algorithms. In addition, future research should also compare the results of different actigraphy devices to determine, in unison with analytics, the most accurate actigraphy and accelerometer devices.

Overall, the results of this study provide significant support that traditional actigraphy algorithms can, with acceptable accuracy, detect sleep and wake in large, diverse population samples, including older adults or populations at risk of health conditions. This study provides researchers and clinicians with evidence that traditional algorithms can continue to be used to assess sleep-wake activity. However, traditional algorithms may present significant limitations in measurement precision of extreme sleep cases. Further, rescoring may not be a necessary step in the analysis of actigraphy data. The implications of this finding are highly important as researchers develop new algorithms and methods of actigraphy data analysis. Specifically, new algorithms should consider the variability of sleep and wake data in sleep problems or extreme cases to precisely measure sleep and wake activity. Further, the results denoted by the current study serve as an important reference point for new, developing strategies for actigraphy analyses.

In conclusion, we emphasize the future direction for generalizable, accurate, and comprehensive actigraphy analyses through the application of machine learning and artificial intelligence models. As aforementioned, several researchers have now adapted artificial intelligence–based models and automated toolboxes to analyze actigraphy and reportedly extract more accurate and more detailed sleep-wake results [53-57]. Given the promise of these methods, we theorize that their assessment and usage to analyze actigraphy data in novel ways poses a promising future direction for more accurate and diverse actigraphy analyses for sleep assessment. Further, in unison with demographic analyses, first of their performance, then the usage of demographic and sleep problem data, we propose this would be a novel opportunity to develop more robust and accurate methods of sleep-wake prediction from activity and actigraphy data. In addition, novel methods using unsupervised machine learning may also provide new ways in which sleep-wake actigraphy data could be parsed to reveal underlying patterns of sleep and sleep problems [56,58]. As these methods are fitted to specific datasets, they could provide better generalizability for sleep-wake analyses [56,58].