Selected Publications
Contact me (link) if you want the pdfs or just check my Google Scholar (link) .
A novel approach for phenotypic characterization of sleep disorders (2024)
Authors: Erhan Bilal, Matheus Lima Diniz Araujo, Catherine M Heinzinger, Samer Ghosn, Nancy Foldvary Schaefer, Carl Y Saab, Reena Mehra, Jeffrey L Rogers
Authors: Erhan Bilal, Matheus Lima Diniz Araujo, Catherine M Heinzinger, Samer Ghosn, Nancy Foldvary Schaefer, Carl Y Saab, Reena Mehra, Jeffrey L Rogers
Obstructive sleep apnea (OSA) is associated with increased morbidity and mortality, afflicting nearly one billion individuals worldwide. Polysomnography is the current gold standard for diagnosing OSA. It is an expensive, timeconsuming exam that relies on a single numerical measure of the frequency of apneas and hypopneas per hour (AHI) to diagnose severity. Considering the rich physiological data collected in each overnight PSG, we foresee a large potential of using automated PSG analysis, leading to a more efficient and informative characterization of OSA pathophysiology. Therefore, we built a real-world dataset (n=10,000) from historical PSG registries, incorporating diverse physiological signals (EEG, ECG, airflow, oximetry, CO2, thoracoabdominal effort) and electronic health records. We implemented strategies to ensure data quality and representation, considering age distribution, sex balance, and inclusion of under-represented minorities, facilitating the conduct of quasi-experimental research within this dataset. Challenges in data curation included artifact identification, historical variability in scoring, and the range of diagnoses in the large cohort. A comparative analysis with other existing PSG study datasets highlighted significant discrepancies, including variations in demographic distributions, comorbidity profiles, longitudinal data continuity, and AHI. These differences emphasize the value of real-world clinical data in providing a comprehensive representation of patient conditions and outcomes, thereby offering deeper insights into the clinical epidemiology of sleep disorders. We anticipate using time series foundational models for the embedding generation and implementing new clinically meaningful metrics beyond AHI. Our proposed approach for the analysis of PSG studies is an unparalleled opportunity to create models that better decipher human physiological patterns, with potential applications extending sleep medicine to disciplines such as neurology, pulmonary, psychiatry, and cardiology.
Excessive daytime sleepiness (EDS) causes difculty in concentrating and continuous fatigue during the day. In the clinical setting, the assessment and diagnosis of EDS rely mostly on subjective questionnaires and verbal reports, which compromises the reliability of clinical diagnosis and the ability to robustly discern candidacy for available therapies and track treatment response. In this study, we used a computational pipeline for the automated, rapid, high-throughput, and objective analysis of previously collected encephalography (EEG) data to identify surrogate biomarkers for EDS, thereby defning the quantitative EEG changes in individuals with high Epworth Sleepiness Scale (ESS) (n= 31), compared to a group of individuals with low ESS (n= 41) at the Cleveland Clinic. The epochs of EEG analyzed were extracted from a large overnight polysomnogram registry during the most proximate period of wakefulness. Signal processing of EEG showed signifcantly diferent EEG features in the low ESS group compared to high ESS, including enhanced power in the alpha and beta bands and attenuation in the delta and theta bands. Our machine learning (ML) algorithms trained on the binary classifcation of high vs. low ESS reached an accuracy of 80.2%, precision of 79.2%, recall of 73.8% and specifcity of 85.3%. Moreover, we ruled out the efects of confounding clinical variables by evaluating the statistical contribution of these variables on our ML models. These results indicate that EEG data contain information in the form of rhythmic activity that could be leveraged for the quantitative assessment of EDS using ML.
Study Objectives
Upper airway stimulation (UAS) therapy is effective for a subset of obstructive sleep apnea (OSA) patients with continuous positive airway pressure (CPAP) intolerance. While overall adherence is high, some patients have suboptimal adherence, which limits efficacy. Our goal was to identify therapy usage patterns during the first three months of therapy to enable targeted strategies for improved adherence.
Methods
Therapy data was retrieved from 2,098 patients for three months after device activation. Data included mean and standard deviation (SD) of hours of use, therapy pauses, hours from midnight the therapy was turned ON and OFF, percentage of missing days, and stimulation amplitude. Cluster analysis was performed using Gaussian mixture models that categorized patients into six main groups.
Results
The six groups and their prevalence can be summarized as Cluster 1A: Excellent Use (34%); Cluster 1B: Excellent Use with variable timing (23%); Cluster 2A: Good Use with missing days and late therapy ON (16%), Cluster 2B: Good Use with missing days, late therapy ON, and early therapy OFF (12%); Cluster 3A: Variable Use with frequent missing days (8%); Cluster 3B: Variable Use with frequent pauses (7%). Most patients (85%) are excellent or good users with mean therapy use >6 hours per night.
Conclusions
Cluster analysis of early UAS usage patterns identified six distinct groups that may enable personalized interventions for improved long-term management. Differentiation of the patient clusters may have clinical implications with regard to sleep hygiene education, therapy discomfort, comorbid insomnia, and other conditions that impact adherence.
Keywords: Cluster Analysis, Obstructive Sleep Apnea, Upper Airway Stimulation, Hypoglossal Nerve Stimulation
Studies in Health Technology and Informatics, 2021
Recombinant human growth hormone (r-hGH) is an established therapy for growth hormone deficiency (GHD); yet, some patients fail to achieve their full height potential, with poor adherence and persistence with the prescribed regimen often a contributing factor. A data-driven clinical decision support system based on “traffic light” visualizations for adherence risk management of patients receiving rhGH treatment was developed. This research was feasible thanks to data-sharing agreements that allowed the creation of these models using real-world data of r-hGH adherence from easypod™ connect; data was retrieved for 11,015 children receiving r-hGH therapy for ≥180 days. Patients’ adherence to therapy was represented using four values (mean and standard deviation [SD] of daily adherence and hours to next injection). Cluster analysis was used to categorize adherence patterns using a Gaussian mixture model. Following a traffic lights-inspired visualization approach, the algorithm was set to generate three clusters: green, yellow, or red status, corresponding to high, medium, and low adherence, respectively. The area under the receiver operating characteristic curve (AUC-ROC) was used to find optimum thresholds for independent traffic lights according to each metric. The most appropriate traffic light used the SD of the hours to the next injection, with an AUCROC value of 0.85 when compared to the complex clustering algorithm. For the daily adherence-based traffic lights, optimum thresholds were >0.82 (SD, <0.37), 0.53–0.82 (SD, 0.37–0.61), and <0.53 (SD, >0.61) for high, medium, and low adherence, respectively. For hours to next injection, the corresponding optimum thresholds were <27.18 (SD, <10.06), 27.18–34.01 (SD, 10.06–29.63), and >34.01 (SD, >29.63). Our research indicates that implementation of a practical data-driven alert system based on recognised traffic-light coding would enable healthcare practitioners to monitor sub-optimally-adherent patients to r-hGH treatment for early intervention to improve treatment outcomes.
Keywords: Adherence, recombinant human growth hormone, growth hormone deficiency, cluster modeling, pediatrics
2021 IEEE 34th International Symposium on Computer-Based Medical Systems (CBMS)
Sleep apnea is a common sleep disorder that, if left untreated, can have critical complications to the individual. The most common and effective treatment for sleep apnea is the Continuous Positive Airway Pressure (CPAP) therapy. But it has a long-term adherence rate as low as 60% due to discomfort and other factors. Although previous research has attempted to increase CPAP usage, there has been little to no change in its average adherence for the past two decades. This paper attempts to change this scenario using a large longitudinal dataset combined with a Recurrent Neural Network model to generate therapy use recommendations after one month of therapy. We performed a retrospective cohort analysis on 3380 patients during their first six months of therapy and compared our personalized recommendation system with the current generic recommendations made by sleep physicians. We show that recommendations generated by our artificial neural network model are easier to achieve since they are significantly closer to patients' therapy progress while being equally successful in maintaining therapy adherence.
Keywords: Recommendation Systems, Therapy Adherence, Recurrent Neural Network, Sleep Apnea, Sleep Medicine
2021 IEEE 34th International Symposium on Computer-Based Medical Systems (CBMS)
Obstructive Sleep Apnea (OSA) is a disorder in which breathing repeatedly stops and starts due to recurrent episodes of partial and complete airway obstruction during sleep. One of the common treatments for moderate and severe OSA cases is the use of Continuous Positive Airway Pressure (CPAP) devices that keep the airways open. Unfortunately, about 40% of the patients using CPAP devices abandon their therapy within six months. In this work, we propose a method to cluster and monitor patients according to their therapy usage behavior aiming for a timely and appropriate intervention. Our data corresponds to 1815 CPAP users in their first six months of therapy. In contrast to the simple rule-based methods currently employed by sleep clinics to identify non-adherent behavior, our approach uses clustering techniques to group patients based on their CPAP usage patterns. After identifying four main clusters, we investigate how patients can change between clusters over the months using Markov Chain analysis. We observed that patients who change to a healthy cluster have a higher probability of staying there in the future, reinforcing the need for early intervention. Finally, we use machine learning-based models to predict the next month's probability of adherence and nonadherence according to our pre-defined cluster definitions.
Keywords: Clustering, Obstructive Sleep Apnea, Sleep Therapy, Therapy Adherence, Markov Chain, Random Forest, SVM, XGBoost, Machine Learning
2021 31st Medical Informatics Europe Conference
The problem of consistent therapy adherence is a current challenge for health informatics, and its solution can increase the success rate of treatments. Here we show a methodology to predict, at individual-level, future therapy adherence for patients receiving daily injections of growth hormone (GH) therapy for GH deficiency. Our proposed model is able to generate predictions of future adherence using a recurrent neural network with adherence data recorded by easypodTM, a connected auto injection device. The model was trained with a multi-year long dataset with 2500 patients, from January 2007 to June 2019. When testing, the model reached an average sensitivity of 0.70 and a specificity of 0.88 per patient when predicting non-adherence (<85%) periods. When evaluated with thousands of therapy segments extracted from a test set, our model reached an AUC-PR score of 0.79 and AUC-ROC of 0.90; both metrics were consistently better than traditional approaches, such as simple average model. Using this model, we can perform precise early identification of patients who are likely to become non-adherent patients. This opens a path for healthcare practitioners to personalize GH therapy at any stage of the patients' journey and improve shared decision making with patients and caregivers to achieve optimal outcomes.
Keywords: Deep learning; growth hormone therapy; therapy adherence.
2019 IEEE International Conference on Big Data (Big Data)
The abandonment rate of patients who use CPAP devices for obstructive sleep apnea (OSA) therapy is as high as 60%. However, there is growing evidence that timely and appropriate intervention can improve long-term adherence to therapy. Current practice in sleep clinics of identifying potential patients who will abandon the treatment is not sufficiently effective in terms of accuracy and timeliness. Recent proposals in the literature have tried to identify non-adherent patients in a specific period of their therapy; however, there is no generalized approach by which clinical providers can monitor their patients continually with the goal of maximizing adherence. Towards this more generic goal, we propose CTAP-CPAP, a Continuous Treatment Adherence Prediction framework. With CTAP-CPAP, we address the problem of generalizing the prediction for any day in the treatment, where a robust framework with multiple machine learning models is implemented to assist medical practitioners keep track of the patient risk of non-adherence. Aiming the parallel progress of both machine learning and health informatics fields, we complement the study with a transparent discussion on the machine learning techniques used to build CTAP-CPAP and our view of its operationalization in a sleep clinic.
Design of Medical Devices Conference, 2020
Upper-Airway Stimulation (UAS) therapy is an innovative alternative to Continuous Positive Airway Pressure (CPAP) treatment for patients with obstructive sleep apnea (OSA) and CPAP intolerance. Patients who have implanted a UAS device are responsible for activating and managing the therapy at home before sleep. Consistent nightly use is required for a reduced OSA burden, measured by the apnea-hypopnea index. Thus, understanding patient behavior and possible challenges to nightly use are crucial to therapy success. In this work, we present two novel visualizations to monitor telemetry data recorded by the UAS sleep remote. They provide doctors and sleep clinicians with detailed information to easily classify therapy use and sleep patterns. We also present how to show daily metrics such as hours of average usage, therapy intensity, and duration of therapy pauses, to identify optimal therapy settings and measure the long-term effectiveness of interventions.
ICWSM 2017
We present an interactive tool that visualizes data on aware-ness of several health conditions in the Middle East. The un-derlying data is obtained via Facebook’s Marketing API and includes rich demographic details. We discuss how this tool may be useful for planning more targeted public health cam-paigns and for monitoring campaign effectiveness.Application URL:http://scdev5.qcri.org/sha/
WebSci, 2017
Every day, millions of users reveal their interests on Facebook, which are then monetized via targeted advertisement marketing campaigns. In this paper, we explore the use of demographically rich Facebook Ads audience estimates for tracking non-communicable diseases around the world. Across 47 countries, we compute the audiences of marker interests, and evaluate their potential in tracking health conditions associated with tobacco use, obesity, and diabetes, compared to the performance of placebo interests. Despite its huge potential, we find that, for modeling prevalence of health conditions across countries, differences in these interest audiences are only weakly indicative of the corresponding prevalence rates. Within the countries, however, our approach provides interesting insights on trends of health awareness across demographic groups. Finally, we provide a temporal error analysis to expose the potential pitfalls of using Facebook's Marketing API as a black box.
EPJ Data Science , 2016
Abstract
In the last few years thousands of scientific papers have investigated sentiment analysis, several startups that measure opinions on real data have emerged and a number of innovative products related to this theme have been developed. There are multiple methods for measuring sentiments, including lexical-based and supervised machine learning methods. Despite the vast interest on the theme and wide popularity of some methods, it is unclear which one is better for identifying the polarity (i.e., positive or negative) of a message. Accordingly, there is a strong need to conduct a thorough apple-to-apple comparison of sentiment analysis methods, as they are used in practice, across multiple datasets originated from different data sources. Such a comparison is key for understanding the potential limitations, advantages, and disadvantages of popular methods. This article aims at filling this gap by presenting a benchmark comparison of twenty-four popular sentiment analysis methods (which we call the state-of-the-practice methods). Our evaluation is based on a benchmark of eighteen labeled datasets, covering messages posted on social networks, movie and product reviews, as well as opinions and comments in news articles. Our results highlight the extent to which the prediction performance of these methods varies considerably across datasets. Aiming at boosting the development of this research area, we open the methods’ codes and datasets used in this article, deploying them in a benchmark system, which provides an open API for accessing and comparing sentence-level sentiment analysis methods.
COSN 2013
Abstract
Several messages express opinions about events, products, and services, political views or even their author’s emotional state and mood. Sentiment analysis has been used in several applications including analysis of the repercussions of events in social networks, analysis of opinions about products and services, and simply to better understand aspects of social communication in Online Social Networks (OSNs). There are multiple methods for measuring sentiments, including lexical-based approaches and supervised machine learning methods. Despite the wide use and popularity of some methods, it is unclear which method is better for identifying the polarity (i.e., positive or negative) of a message as the current literature does not provide a method of comparison among existing methods. Such a comparison is crucial for understanding the potential limitations, advantages, and disadvantages of popular methods in analyzing the content of OSNs messages. Our study aims at filling this gap by presenting comparisons of eight popular sentiment analysis methods in terms of coverage (i.e., the fraction of messages whose sentiment is identified) and agreement (i.e., the fraction of identified sentiments that are in tune with ground truth). We develop a new method that combines existing approaches, providing the best coverage results and competitive agreement. We also present a free Web service called iFeel, which provides an open API for accessing and comparing results across different sentiment methods for a given text.
SAC 2016
Abstract
Sentiment analysis has become a key tool for several social media applications, including analysis of user’s opinions about products and services, support to politics during campaigns and even for market trending. There are multiple existing sentiment analysis methods that explore different techniques, usually relying on lexical resources or learning approaches. Despite the large interest on this theme and amount of research efforts in the field, almost all existing methods are designed to work with only English content. Most existing strategies in specific languages consist of adapting existing lexical resources, without presenting proper validations and basic baseline comparisons. In this paper, we take a different step into this field. We focus on evaluating existing efforts proposed to do language specific sentiment analysis. To do it, we evaluated twenty-one methods for sentence-level sentiment analysis proposed for English, comparing them with two language-specific methods. Based on nine language-specific datasets, we provide an extensive quantitative analysis of existing multi-language approaches. Our main result suggests that simply translating the input text on a specific language to English and then using one of the existing English methods can be better than the existing language specific efforts evaluated. We also rank those implementations comparing their prediction performance and identifying the methods that acquired the best results using machine translation across different languages. As a final contribution to the research community, we release our codes and datasets. We hope our effort can help sentiment analysis to become English independent.
WWW 2014
Abstract
Sentiment analysis methods are used to detect polarity in thoughts and opinions of users in online social media. As businesses and companies are interested in knowing how social media users perceive their brands, sentiment analysis can help better evaluate their product and advertisement campaigns. In this paper, we present iFeel, a Web application that allows one to detect sentiments in any form of text including unstructured social media data. iFeel is free and gives access to seven existing sentiment analysis methods: SentiWordNet, Emoticons, PANAS-t, SASA, Happiness Index, SenticNet, and SentiStrength. With iFeel, users can also combine these methods and create a new Combined-Method that achieves high coverage and F-measure. iFeel provides a single platform to compare the strengths and weaknesses of various sentiment analysis methods with a user friendly interface such as file uploading, graphical visualizing, and weight tuning.