A Personalized Adaptive Algorithm for Sleep Quality Prediction Using Physiological and Environmental Sensing Data A Personalized Adaptive Algorithm for Sleep Quality Prediction using Physiological and[.]
2021 8th NAFOSTED Conference on Information and Computer Science (NICS) A Personalized Adaptive Algorithm for Sleep Quality Prediction using Physiological and Environmental Sensing Data 1st Nguyen Thi Phuoc Van 2nd Dao Minh Son 3rd Koji Zettsu Big data Integration Research Center NICT Tokyo, Japan phuocvan@nict.go.jp Big data Integration Research Center NICT Tokyo, Japan dao@nict.go.jp Big data Integration Research Center NICT Tokyo, Japan zettsu@nict.go.jp Abstract—The lacking data from wearable sensors to solve different problems in the healthcare area is obvious since it is not easy to find enough volunteers to collect data Moreover, human reacts very differently to medical treatment/ exercise levels/ stress and so on Therefore, we need an advanced prediction model which can reuse the public data and can be adapt to personal data to predict health parameters This paper introduces a solution for this issue We present a novel personalized adaptive algorithm based on ensemble learning to predict sleeping efficiency, the proposed framework can be extended to solve many problems in healthcare applications In this work, the global model is built based on ensemble learning with common features from all clients The global model is then combined with the model from the client with more personalized features The client model will learn and be updated model every day Our proposed framework was tested in two data sets PMData and another private data set and showed better results than the conventional method The proposed algorithm/ framework is a great step to solve the prediction problem in healthcare since each person has their own characteristics, responds differently to treatments/environment/stressful levels The proposed algorithm is a big enhancement in building a health navigator system to enhance human health Index Terms—Sleeping efficiency; transfer learning; Random forest; health navigator; activities; adaptive prediction model I I NTRODUCTION LEEPING quality has an important role in human life Poor sleep may lead to many problems like reduction of concentration and productivity, affecting negatively on blood sugar and increasing type diabetes risk, reducing the immune system, or decreasing the performance of social interactions [1]–[13] Since maintaining good sleeping is a crucial factor to be perfectly healthy We need to know which factors cause good sleep, is that possible to predict sleeping quality/ sleep efficiency from people activities, environmental factors, and so on The development of wearable sensors, environmental S 978-1-6654-1001-4/21/$31.00 ©2021 IEEE sensors, embedded systems, and big data makes it possible to collect information to predict sleeping quality and find the reasons to enhance sleep quality The previous prediction models were built based on machine learning techniques and introduced quite good results However, each person reacts differently to the physical activities, environmental parameters, and so on Multiple factors affect sleep quality like burned calories, active levels, stress levels Therefore, necessary to establish a personalized adaptive sleep prediction model for each individual is essential to develop the e-healthcare application Personalized/adaptive models have been utilized in other areas [14]–[16] For the data center resource utilization estimation [15], the adaptive prediction model refers to the selection of the best model in the given time window Another recent paper [16] proposed a model which is adjusted seasonality and removed the error cycle Aside from personalized adaptive for each user, the model could be more reliable if it gets knowledge from a large mass database - this idea leads to the using of transfer learning in sleep science [17], [18] Phan et al [18] used transfer learning for sleep staging and showed the impressive improvement of a prediction model The idea of using transfer learning is a good foundation to build the adaptive model in sleep quality prediction Even though the personalized adaptive models are applied in many fields but not much in sleep research Transfer learning was also used to improve sleep stages classification Besides sleep stages, sleep efficiency is one of the substantial sleep attributes sleep efficiency is the proportion between sleep duration and total time spent on the bed Currently, we not have personalized adaptive models to predict sleep efficiency [19] For this reason, this work investigates a personalized 113 2021 8th NAFOSTED Conference on Information and Computer Science (NICS) adaptive algorithm for sleep efficiency from activities, environmental factors and compares it to a recent popular model Our proposed prediction model gives better results than the transfer deep learning models and Bayesian Replicator Neural Network on the public and data sets Our proposed framework fosters sleep quality prediction in the e-healthcare system This method is reusable in different applications by using a trained model on the public data on private data event private data has more features to improve the prediction accuracy The contribution of this study is as follows: • We introduce a program/algorithm to pre-process public data for sleep quality prediction problems This program created a data frame that contains features and sleep attributes The data frame can be used to find correlation between activities and sleep attributes and to build prediction model for sleep efficiency • The personalized adaptive model to predict sleep efficiency was developed and shows the great capability in reusing public data to solve the accuracy modeling problem in the private data Our framework is utilized to solve the lacking data issue in building the model in the healthcare area II R ELATED WORK A Association between sleep quality and environmental factors Several studies focus on the finding association rules between environment and sleep duration [20]–[23] They considered the relationship between the natural/ social environment and sleep quality In these kinds of studies, the sleep quality is evaluated by questionnaires- Pittsburgh Sleep Quality Index (PSQI) Reference [20] found the relation between environment and sleep quality in the summer of Japan This research showed that the air-condition and light conditions might affect the sleep of volunteers Controlling the room temperature and light can improve sleep quality More broadly considered, the study [21] considered the relationship between air pollution and sleep duration in young people in Beijing China The sleep duration was collected by doing the sleep survey- Chinese version P SQI metric The data of sleep and environment were collected in the five-year duration (2013 − 2018) from 16, 889 students This work concluded that sleep duration reduction has a high association with air pollution Another study [22] found the sleep disorder in old people and the environment This research utilized the health database of Chinese elderly in Ningbo province in the duration from 2008 to 2017 The data information of visiting the hospital of old people at the age (60+) was collected along with the daily air pollution parameters like nitrogen dioxide, sulfur dioxide, inhale-able particles, and so forth This work discovered that the air pollution exposure in the old people had a high association with the frequency of doctor/clinic visits to solve the sleep disorder problems To find association rules between sleep quality/duration and environmental factors, in almost all studies, the questionnaire - P SQI was used as a sleep evaluation metric However, this kind of evaluation is not always accurate since answers are subjective Moreover, sometimes, people forget to answer the question at the right time and even forget the feeling of themself after sleep B Association between sleep quality and daily activities The development of wearable sensors makes it easy for us to collect stages of sleeping and obtain more accurate sleeping quality metrics Moreover, physical activities also can be collected by wearable sensors [24], [25] Aarti Sathyanarayana et al [2] develop a deep learning-based model to predict sleeping quality (good/poor) from tracking physical activities in the awake time Their prediction model showed a good result (up to 94 %), much higher than the logistic regression model The physiological signals are also used to predict sleeping quality in the caregivers of people with dementia (CP W D) group [26] The wearable sensor (E4 Empatica wristband) was used to collect body movement (through accelerometer), heart rate, electrodermal activity, and skin temperature Those features are used to measure sleeping quality and restfulness in CP W D and yielded a good prediction result - up to 75% III PERSONALIZED A DAPTIVE ALGORITHM The objectives of this method are (1) to choose the best global model which is generated from common features of public data, and (2) to combine the global model and personal model with extra private features to build the personalized adaptive model for each individual The random forest regression based on the least-squares error criterion is used in the global model for its flexibility, easy comprehensibility, and computational efficiency The tree-based regression model provides the propositional logic representation of predictor space in a tree form A logical test on a predictor variable is done in each internal node of a tree The framework is started with the constructed with the global model This model is then distributed to each client to establish the client model The proposed framework is illustrated in Fig In this work, the global model is built from global data In the global data, only common features of all clients are used to train the model Global data is divided into two parts, 75% for training and 25% for testing We utilized the stacking ensemble learning in the global model to get better performance of the model After checking the model on the 114 2021 8th NAFOSTED Conference on Information and Computer Science (NICS) testing data, if the model has good performance it will be sent to each client to reuse The algorithm of the global model is described in III-A At the client, the global is combined to the other model which is shaped from client features The details of the algorithm at each client is discussed in the III-B In each node k the error (M SE) is calculated as follows [27]: !2 k k X X M SEk = yi − yi (2) nk nk where nk is the number samples of node k The error of a tree is defined as the average of errors in its leaves The tree is developed based on two sets of splitting rules, the goal of building the tree/splitting rule is to minimize the error of this tree This rule is used to build model and model - two main components to build the global model The algorithm of the global model is presented as follows: B Personalized Adaptive client model This model is built from global model and personal model based on the personal data The detail of algorithm is presented in Algorithm The first step is to load the global model, after that, based on number of important features in the client we choose the structure of local model This model is then combined with global model by stacking technique to make personalized Adaptive model (M odelCA ) Fig Proposed personalized adaptive framework IV DATA SETS A PMData dataset A Proposed global model The global model is based on ensemble learning to combine two models into one to optimize the prediction results Model and model are Random Forest Regression models The optimized tree of these models was calculated by minimizing the Mean Square Error (M SE) ns X (yi − ypre (xi )) ns i (1) where ns is the number of data samples; yi is output value i, ypre (xi ) is the predicted value of the output based on the set of feature x order i An example of a brand of tree based model is described in PMData data set (https : //osf.io/vx4bk/wiki/home/) was collected by the Fitbit Versa smartwatch, PMSys sports logging, and google form for the duration of 150 days of 16 participants In total, we have 1663 days of data Data from Fitbit consists of information about burned calories/ min, moved distance/min, heartbeats/min, steps/min, time in different heartbeat zone, light/moderate/very active minutes per day, sleep score, report about meals, report about wellness (fatigue, mood, readiness ) Most of files are formatted as *.json and *.csv This data was used for competition and showed the possibility for analysis In this work, the activities information like distances, sedentary, steps, sleeping is used to validate our proposed sleeping efficiency prediction model The process diagram to collect and gather PMData is showed in the Fig and the algorithm is shown in We can see in Fig that the extra features including heartbeat (average heartbeat of the person in a day), pm sise (the size of particulars in the air), humidity, pm10 (particulate matter lower than 10µm), temperature, CO2 (level of carbon dioxide) are considered to build the adaptive model for each person B Fitness data set Fig An example of brand of a random forest regression This data consists of sub-set data Each sub-set data has information about the activities and environment of a volunteer in the duration of 90 days The ethics of the data collection campaign was approved by The National Institute 115 2021 8th NAFOSTED Conference on Information and Computer Science (NICS) Algorithm 1: Proposed global model Input: Input: Data consists of all common features Output: Output: Global model Initialization; 1: Select group1 of features 2: Select group2 of features 3: Build model Input1: set of training data (ntrain data points) group1 with c features, {hxtraini , ytraini i} , i = 1, 2, ntrain ; xtraini = {f eature0 , f eature1 , , f eaturec } Output1: a regression prediction model Set the initial tree structure T ∗ =< initialsplit > For all possible tree Tv Calculate M SE(Tv ) Find the Tbest where M SE(Tbest ) = minimum value of M SE(Tv ) end For If M SE(Tbest ) > M SE(T ∗ ) Create model1 with T ∗ else Create model1 with Tbest 4: Build model Input2: set of training data (ntrain data points) group1 with d features (d > c), {hxtraindi , ytraini i} , i = 1, 2, ntrain ; xtraindi = {f eaturec+1 , f eaturec+2 , , f eatured } Output2: a regression prediction model Set the initial tree structure T d∗ =< initialsplit > For all possible tree Tv Calculate M SE(T dv ) Find the T dbest where M SE(T dbest ) = minimum value of M SE(T dv ) end For If M SE(T dbest ) > M SE(T d∗ ) Create model2 with T d∗ else Create model2 with T dbest 5: Combine model1 and model to a global model 6: Check performance of global on testing data 7: Save global model Return: Output: Global model Algorithm 2: Proposed client model Input: Global model, Client data consists of all common features and personal features Output: Client prediction model Initialization; 1: Load global model 2: Choose the structure of client model 3: Set the initial training data set Xtrain , ytrain 4: Build the client model in the similar way of model1 in the global algorithm 5: Combine global model and client model to make the personalized Adaptive model (M odelCA ) 6: Predict the next outcome based on previous data For i in testing data ypredictedi+1 = apply (M odelCA ) on data set Xtrain ∪ Xi Error(i + 1) = ypredictedi+1 − ylabeli+1 end For 7: Evaluate the performance of adaptive model Calculate Root Mean Square Error Calculate Mean Absolute Error Return: Output: Client model Fig Phases to collect and preprocess PMData of Information and Communications Technology, Japan Each person wears a smartwatch Fitbit sense to collect activities and sleeping data Volunteers were provided smartphones to answer the questionnaire and the environmental sensors were installed in their bedroom to collect surrounding data (noise, P M 25, temperature, CO2, P M size, humidity) In total, 783 days of data were collected The data from Fitbit sense was 116 2021 8th NAFOSTED Conference on Information and Computer Science (NICS) used to validate the proposed framework Several state-of-art models mentioned in [18], [28] also applied, in the same manner, to compare with the proposed one Two deep transfer learning models with two hidden layers with shape(30, 3, 1) and shape(30, 4, 1) are reviewed to compare with the proposed model In addition, we also make trial run on the Bayesian Replicator Neural Network [28] and Tree-based models which are popular for small data sets A popular accuracy metrics - Root Mean Square Error (RM SE) is used to compare the error prediction between models Figure shows that in all cases Algorithm 3: Pmdata- Data prepossessing steps Require: Input: RaW data from Fitbit, PMSys, app, and google form Ensure: Output: processed Data Initialization; 1: Define the function dateo nly(time) Calculate the date of date and time string return date string 2: Define the function loadjs(∗.json) load *.json file return content of json file in dictionary format 3: Define the function loada llj sonf it(path) Use loadjs function to load data (∗.json files) in the Fitbit folder then convert data into data frame 4: Load and change sleep.json to dataframe by loadjs function and pd.DataFrame 5: Load data from exercise.json by loadjs function 6: Load data from *.json file Calculate feature values for each day Create a dataframe has information of date and feature values 7: Load data from *.csv file 8: Merge all above data to a data frame named based on the common date, the data of missing days will be removed 9: return alldata 10: end Fig Compare the RMSE of proposed model and recent models on PMData processed in a similar of processing PMData This data is used to validate the personalized adaptive model again to show the performance of our model V R ESULTS AND DISCUSSION In this section, the results of the proposed model are taken into account and compared with the conventional one For both data set, we applied Leave One Out (LOO) technique to check the performance of the proposed model A Results on PMData data set After cleaning and processing PMData, data of 15 participants was used to build and validate the model.150-day activities data of each person is considered as a subset Data of 14 people with common features (very active minutes, moderate active minutes, light active minutes, sedentary minutes, distance, steps ) was used to form a global model to predict sleeping efficiency (ranged 61-100, mean value = 87.8, standard deviation = 3.7) The rest data (1 person) with 10 features (6 common features in the global model and four extra features are time in four zones of heartbeat) was the proposed framework has higher prediction accuracy/less error The average RM SE of our framework is 3.5 while the RM SE of other prediction methods are larger than For example, at users ID 10 the different error between the Treebased transfer learning and personalized adaptive model is 1.7 ( around 40%) To check the practical application of our proposed framework, we check its performance on another data set which was collected in different locations and participants did various types of exercises In the next section, the RM SE of some state-of-art models and proposed model on the fitness data are presented and compared B Results on group fitness data Steps to reuse public data to build the personalized adaptive prediction model for personal sleeping efficiency prediction are demonstrated in Figure The PMData is used to create the global model activity features in PMData are reused to construct the global model, more environmental parameters in the client are added to create the personalized adaptive model Figure The performance of the proposed frame work is examined through RM SE Figure reveals again the superiority of the adaptive model Our proposed model has lowest error The average error on our model decrease 25% and 50%in comparison with deep transfer learning and Bayesian RNN 117 2021 8th NAFOSTED Conference on Information and Computer Science (NICS) C Discussion It is obvious that the healthcare prediction model problem always faces the lacking of data since the number of volunteers is not large enough or missing the diversity of participants in gender, age range, health condition, and so on In addition, each person responds very differently to the level of exercise practicing, weather, air pollution, medical treatment, and so on Our personalized adaptive prediction for sleeping efficiency demonstrates the first important point that the prediction model in healthcare applications should be built on the personal level with data over a long period The second good point in our framework is to reuse the public data to increase the reliability and accuracy of the personal model This point is proofed in the observation of the RM SE values of public data and private data, the private data shows less error VI C ONCLUSION Fig Steps to apply personalized adaptive framework on fitness group data Our proposed algorithm shows a high accuracy prediction in comparison with the traditional one We use the sleeping efficiency prediction from activities and the environment to demonstrate the advantages of our proposed model Our framework opens a good approach to solve the modeling issue in the healthcare area This framework confirms the possibility of using public data (large data set) for private applications with small data set but compatible features Moreover, this framework hints at the use of federated learning that all clients contribute common features to build the global model Each client then can use a global model to create its model ACKNOWLEDGMENT We would like to thank GreenBlue and TAOS for sharing materials We also would like to thank for the contribution of DLC to help us to collect data for this research R EFERENCES Fig Compare the RMSE of proposed model and state-of-art models on group fitness data model, respectively Moreover, this error in this private data yields the better result than in public data since in private data we obtained more feature from environment, those features (like noise, temperature) have high impact on sleeping efficiency of people 118 [1] Reza Boostani, Foroozan Karimzadeh, and Mohammad Nami, “A comparative review on sleep stage classification methods in patients and healthy individuals,” Computer methods and programs in biomedicine, vol 140, pp 77–91, 2017 [2] Aarti Sathyanarayana, Shafiq Joty, Luis Fernandez-Luque, Ferda Ofli, Jaideep Srivastava, Ahmed Elmagarmid, Teresa Arora, and Shahrad Taheri, “Sleep quality prediction from wearable data using deep learning,” JMIR mHealth and uHealth, vol 4, no 4, pp e125, 2016 [3] Joao Palotti, Raghvendra Mall, Michael Aupetit, Michael Rueschman, Meghna Singh, Aarti Sathyanarayana, Shahrad Taheri, and Luis Fernandez-Luque, “Benchmark on a large cohort for sleep-wake classification with machine learning techniques,” NPJ digital medicine, vol 2, no 1, pp 1–9, 2019 https://greenblue.co.jp/ http://www.taos.tokyo/ 2021 8th NAFOSTED Conference on Information and Computer Science (NICS) [4] F´abio Mendonc¸a, Sheikh Shanawaz Mostafa, Fernando Morgado-Dias, Antonio G Ravelo-Garcia, and Thomas Penzel, “A review of approaches for sleep quality analysis,” Ieee Access, vol 7, pp 24527–24546, 2019 [5] Junming Zhang and Yan Wu, “A new method for automatic sleep stage classification,” IEEE transactions on biomedical circuits and systems, vol 11, no 5, pp 1097–1110, 2017 [6] Emina Alickovic and Abdulhamit Subasi, “Ensemble svm method for automatic sleep stage classification,” IEEE Transactions on Instrumentation and Measurement, vol 67, no 6, pp 1258–1265, 2018 [7] Ch Panagiotou, I Samaras, J Gialelis, Petros Chondros, and Dimitrios Karadimas, “A comparative study between svm and fuzzy inference system for the automatic prediction of sleep stages and the assessment of sleep quality,” in 2015 9th International Conference on Pervasive Computing Technologies for Healthcare (PervasiveHealth) IEEE, 2015, pp 293–296 [8] Helen Skouteris, Carmela Germano, Eleanor H Wertheim, Susan J Paxton, and Jeannette Milgrom, “Sleep quality and depression during pregnancy: a prospective study,” Journal of sleep research, vol 17, no 2, pp 217–220, 2008 [9] Ying Yang, Jing Mao, Zhiying Ye, Xiaoli Zeng, Huimin Zhao, Yueting Liu, and Jie Li, “Determinants of sleep quality among pregnant women in china: a cross-sectional survey,” The Journal of Maternal-Fetal & Neonatal Medicine, vol 31, no 22, pp 2980–2985, 2018 [10] Pham Son Lam, Nguyen Dinh Son, Hoang Phuong Chi, Nguyen Thi Phuoc Van, and Nguyen Duc Minh, “Novel algorithm to classify sleep stages,” in 2019 13th International Conference on Sensing Technology (ICST), 2019, pp 1–6 [11] Yan-Ping Bao, Ying Han, Jun Ma, Ru-Jia Wang, Le Shi, Tong-Yu Wang, Jia He, Jing-Li Yue, Jie Shi, Xiang-Dong Tang, et al., “Cooccurrence and bidirectional prediction of sleep disturbances and depression in older adults: meta-analysis and systematic review,” Neuroscience & Biobehavioral Reviews, vol 75, pp 257–273, 2017 [12] David G Mcsharry, Silke Ryan, Peter Calverley, J Colin Edwards, and Walter T Mcnicholas, “Sleep quality in chronic obstructive pulmonary disease,” Respirology, vol 17, no 7, pp 1119–1124, 2012 [13] Thi Phuoc Van Nguyen, Do Van Nguyen, and Koji Zettsu, “Models to predict sleeping quality from activities and environment: Current status, challenges and opportunities,” in Proceedings of the 2021 Workshop on Intelligent Cross-Data Analysis and Retrieval, 2021, pp 52–56 [14] Zhong Yin, Yongxiong Wang, Li Liu, Wei Zhang, and Jianhua Zhang, “Cross-subject eeg feature selection for emotion recognition using transfer recursive feature elimination,” Frontiers in neurorobotics, vol 11, pp 19, 2017 [15] Waheed Iqbal, Josep Lluis Berral, Abdelkarim Erradi, David Carrera, et al., “Adaptive prediction models for data center resources utilization estimation,” IEEE Transactions on Network and Service Management, vol 16, no 4, pp 1681–1693, 2019 [16] Dyego HL Oliveira, MV Francisco Filho, Thelmo P de Ara´ujo, Joaquim Celestino, and Rafael L Gomes, “Adaptive model for network resources prediction in modern internet service providers,” in 2020 IEEE Sym- [17] [18] [19] [20] [21] [22] [23] [24] [25] [26] [27] [28] 119 posium on Computers and Communications (ISCC) IEEE, 2020, pp 1–6 Sara Mahvash Mohammadi, Samaneh Kouchaki, Saeid Sanei, DerkJan Dijk, Adrian Hilton, and Kevin Wells, “Tensor factorisation and transfer learning for sleep pose detection,” in 2019 27th European Signal Processing Conference (EUSIPCO), 2019, pp 1–5 Huy Phan, Oliver Y Ch´en, Philipp Koch, Alfred Mertins, and Maarten De Vos, “Deep transfer learning for single-channel automatic sleep staging with channel mismatch,” in 2019 27th European Signal Processing Conference (EUSIPCO) IEEE, 2019, pp 1–5 Ignacio Perez-Pozuelo, Bing Zhai, Joao Palotti, Raghvendra Mall, Michaăel Aupetit, Juan M Garcia-Gomez, Shahrad Taheri, Yu Guan, and Luis Fernandez-Luque, “The future of sleep health: a data-driven revolution in sleep science and medicine,” NPJ digital medicine, vol 3, no 1, pp 1–15, 2020 Momoko Kayaba, Tomohiko Ihara, Hiroyuki Kusaka, Satoru Iizuka, Kenji Miyamoto, and Yasushi Honda, “Association between sleep and residential environments in the summertime in japan,” Sleep medicine, vol 15, no 5, pp 556–564, 2014 Hongjun Yu, Panpan Chen, Shelby Paige Gordon, Miao Yu, and Yangyang Wang, “The association between air pollution and sleep duration: a cohort study of freshmen at a university in beijing, china,” International journal of environmental research and public health, vol 16, no 18, pp 3362, 2019 Mengling Tang, Die Li, Zeyan Liew, Fang Wei, Jianbing Wang, Mingjuan Jin, Kun Chen, and Beate Ritz, “The association of short-term effects of air pollution and sleep disorders among elderly residents in china,” Science of The Total Environment, vol 708, pp 134846, 2020 Bing Cao, Yan Chen, and Roger S McIntyre, “Comprehensive review of the current literature on impact of ambient air pollution and sleep quality,” Sleep Medicine, vol 79, pp 211–219, 2021 Bao-Lan Le-Quang, Minh-Son Dao, and Mohamed Saleem Haja Nazmudeen, “Wemotion: A system to detect emotion using wristbands and smartphones,” in International Symposium on Signal Processing and Intelligent Recognition Systems Springer, 2018, pp 92–103 Minh-Son Dao and Mohamed Saleem Haja Nazmudeen, “Internet of assistants: Humans can get assistance anywhere, anytime and any areas,” in International Symposium on Signal Processing and Intelligent Recognition Systems Springer, Singapore, 2019, pp 13–22 Reza Sadeghi, Tanvi Banerjee, Jennifer C Hughes, and Larry W Lawhorne, “Sleep quality prediction in caregivers using physiological signals,” Computers in biology and medicine, vol 110, pp 276–288, 2019 Lu´ıs Torgo, “Inductive learning of tree-based regression models,” Ai Communications, vol 13, 04 2000 Iman Azimi, Olugbenga Oti, Sina Labbaf, Hannakaisa Niela-Vilen, Anna Axelin, Nikil Dutt, Pasi Liljeberg, and Amir M Rahmani, “Personalized maternal sleep quality assessment: An objective iot-based longitudinal study,” IEEE Access, vol 7, pp 93433–93447, 2019 ... [2] Aarti Sathyanarayana, Shafiq Joty, Luis Fernandez-Luque, Ferda Ofli, Jaideep Srivastava, Ahmed Elmagarmid, Teresa Arora, and Shahrad Taheri, ? ?Sleep quality prediction from wearable data using. .. dataframe by loadjs function and pd.DataFrame 5: Load data from exercise.json by loadjs function 6: Load data from *.json file Calculate feature values for each day Create a dataframe has information... deep learning,” JMIR mHealth and uHealth, vol 4, no 4, pp e125, 2016 [3] Joao Palotti, Raghvendra Mall, Michael Aupetit, Michael Rueschman, Meghna Singh, Aarti Sathyanarayana, Shahrad Taheri, and