Health data space nodes for privacy-preserving linkage of medical data to support collaborative secondary analyses.

Introduction: The potential for secondary use of health data to improve healthcare is currently not fully exploited. Health data is largely kept in isolated data silos and key infrastructure to aggregate these silos into standardized bodies of knowledge is underdeveloped. We describe the development, implementation, and evaluation of a federated infrastructure to facilitate versatile secondary use of health data based on Health Data Space nodes. Materials and methods: Our proposed nodes are self-contained units that digest data through an extract-transform-load framework that pseudonymizes and links data with privacy-preserving record linkage and harmonizes into a common data model (OMOP CDM). To support collaborative analyses a multi-level feature store is also implemented. A feasibility experiment was conducted to test the infrastructures potential for machine learning operations and deployment of other apps (e.g., visualization). Nodes can be operated in a network at different levels of sharing according to the level of trust within the network. Results: In a proof-of-concept study, a privacy-preserving registry for heart failure patients has been implemented as a real-world showcase for Health Data Space nodes at the highest trust level, linking multiple data sources including (a) electronical medical records from hospitals, (b) patient data from a telemonitoring system, and (c) data from Austria's national register of deaths. The registry is deployed at the tirol kliniken, a hospital carrier in the Austrian state of Tyrol, and currently includes 5,004 patients, with over 2.9 million measurements, over 574,000 observations, more than 63,000 clinical free text notes, and in total over 5.2 million data points. Data curation and harmonization processes are executed semi-automatically at each individual node according to data sharing policies to ensure data sovereignty, scalability, and privacy. As a feasibility test, a natural language processing model for classification of clinical notes was deployed and tested. Discussion: The presented Health Data Space node infrastructure has proven to be practicable in a real-world implementation in a live and productive registry for heart failure. The present work was inspired by the European Health Data Space initiative and its spirit to interconnect health data silos for versatile secondary use of health data.

Neural Network-Based Prediction of Perceived Sleep Quality Through Wearable Device Data.

BACKGROUND: This study focuses on the development of a neural network model to predict perceived sleep quality using data from wearable devices. We collected various physiological metrics from 18 participants over four weeks, including heart rate, physical activity, and both device-measured and self-reported sleep quality. OBJECTIVES: The primary objective was to correlate wearable device data with subjective sleep quality perceptions. METHODS: Our approach used data processing, feature engineering, and optimizing a Multi-Layer Perceptron classifier. RESULTS: Despite comprehensive data analysis and model experimentation, the predictive accuracy for perceived sleep quality was moderate (59%), highlighting the complexities in accurately quantifying subjective sleep experiences through wearable data. Applying a tolerance of 1 grade (on a scale from 1-5), increased accuracy to 92%. DISCUSSION: More in-depth analysis is required to fully comprehend how wearables and artificial intelligence might assist in understanding sleep behavior.

Instance Selection Algorithms for Predictive Modelling in Telehealth Applications.

Telehealth services are becoming more and more popular, leading to an increasing amount of data to be monitored by health professionals. Machine learning can support them in managing these data. Therefore, the right machine learning algorithms need to be applied to the right data. We have implemented and validated different algorithms for selecting optimal time instances from time series data derived from a diabetes telehealth service. Intrinsic, supervised, and unsupervised instance selection algorithms were analysed. Instance selection had a huge impact on the accuracy of our random forest model for dropout prediction. The best results were achieved with a One Class Support Vector Machine, which improved the area under the receiver operating curve of the original algorithm from 69.91 to 75.88 %. We conclude that, although hardly mentioned in telehealth literature so far, instance selection has the potential to significantly improve the accuracy of machine learning algorithms.

Identifying Accuracy-related Outliers in a 6-Minute Walk Test Application for Monitoring Cardiovascular Disease Patients.

The 6-Minute Walk Test (6-MWT) is frequently used to evaluate functional physical capacity of patients with cardiovascular diseases. To determine reliability in remote care, outlier classification of a mobile Global Navigation Satellite System (GNSS) based 6-MWT App had to be investigated. The raw data of 53 measurements were Kalman filtered and afterwards layered with a Butterworth high-pass filter to find correlation between the resulting Root Mean Square value (RMS) outliers to relative walking distance errors using the test. The analysis indicated better performance in noise detection using all 3 GNSS dimensions with a high Pearson correlation of r = 0.77, than sole usage of elevation data with r = 0.62. This approach helps with the identification between accurate and unreliable measurements and opens a path that allows usage of the 6-MWT in remote disease management settings.Clinical Relevance- The 6-MWT is an important assessment tool of walking performance for patients with cardiovascular diseases. Using a sufficiently accurate application would enable unsupervised and easy remote usage, which could potentially reduce the demand for in-clinic visits and facilitate a more convenient and reliable monitoring method in telehealth settings.

Introduction and Comparison of Novel Decentral Learning Schemes with Multiple Data Pools for Privacy-Preserving ECG Classification.

Artificial intelligence and machine learning have led to prominent and spectacular innovations in various scenarios. Application in medicine, however, can be challenging due to privacy concerns and strict legal regulations. Methods that centralize knowledge instead of data could address this issue. In this work, 6 different decentralized machine learning algorithms are applied to 12-lead ECG classification and compared to conventional, centralized machine learning. The results show that state-of-the-art federated learning leads to reasonable losses of classification performance compared to a standard, central model (-0.054 AUROC) while providing a significantly higher level of privacy. A proposed weighted variant of federated learning (-0.049 AUROC) and an ensemble (-0.035 AUROC) outperformed the standard federated learning algorithm. Overall, considering multiple metrics, the novel batch-wise sequential learning scheme performed best (-0.036 AUROC to baseline). Although, the technical aspects of implementing them in a real-world application are to be carefully considered, the described algorithms constitute a way forward towards preserving-preserving AI in medicine.

Effect of digital tools in outpatient cardiac rehabilitation including home training-results of the EPICURE study.

Introduction: Cardiovascular diseases are the leading cause of death worldwide and are partly caused by modifiable risk factors. Cardiac rehabilitation addresses several of these modifiable risk factors, such as physical inactivity and reduced exercise capacity. However, despite its proven short-term merits, long-term adherence to healthy lifestyle changes is disappointing. With regards to exercise training, it has been shown that rehabilitation supplemented by a) home-based exercise training and b) supportive digital tools can improve adherence. Methods: In our multi-center study (ClincalTrials.gov Identifier: NCT04458727), we analyzed the effect of supportive digital tools like digital diaries and/or wearables such as smart watches, activity trackers, etc. on exercise capacity during cardiac rehabilitation. Patients after completion of phase III out-patient cardiac rehabilitation, which included a 3 to 6-months lasting home-training phase, were recruited in five cardiac rehabilitation centers in Austria. Retrospective rehabilitation data were analyzed, and additional data were generated via patient questionnaires. Results: 107 patients who did not use supportive tools and 50 patients using supportive tools were recruited. Already prior to phase III rehabilitation, patients with supportive tools showed higher exercise capacity (Pmax = 186 ± 53 W) as compared to patients without supportive tools (142 ± 41 W, p < 0.001). Both groups improved their Pmax, significantly during phase III rehabilitation, and despite higher baseline Pmax of patients with supportive tools their Pmax improved significantly more (ΔPmax = 19 ± 18 W) than patients without supportive tools (ΔPmax = 9 ± 17 W, p < 0.005). However, after adjusting for baseline differences, the difference in ΔPmax did no longer reach statistical significance. Discussion: Therefore, our data did not support the hypothesis that the additional use of digital tools like digital diaries and/or wearables during home training leads to further improvement in Pmax during and after phase III cardiac rehabilitation. Further studies with larger sample size, follow-up examinations and a randomized, controlled design are required to assess merits of digital interventions during cardiac rehabilitation.

Towards a Verification Approach of a Smart Registry for Chronic Heart Failure Patients.

BACKGROUND: The daily increasing amount of health data from different sources like electronic medical records and telehealth systems go hand in hand with the ongoing development of novel digital and data-driven analytics. Unifying this in a privacy-preserving data aggregation infrastructure can enable services for clinical decision support in personalized patient therapy. OBJECTIVES: The goal of this work was to consider such an infrastructure, implemented in a smart registry for heart failure, as a comparative method for the analysis of health data. METHODS: We analyzed to what extent the dataset of a study on the telehealth program HerzMobil Tirol (HMT) can be reproduced with the data from the smart registry. RESULTS: A table with 96 variables for 251 patients of the HMT publication could theoretically be replicated from the smart registry for 248 patients with 80 variables. The smart registry contained the tables to reproduce a large part of the information, especially the core statements of the HMT publication. CONCLUSION: Our results show how such an infrastructure can enable efficient analysis of health data, and thus take a further step towards personalized health care.

Impact of Professional Background on Inter-Annotator Variability and Accuracy During Annotation of Clinical Notes.

BACKGROUND: The aging population's need for treatment of chronic diseases is exhibiting a marked increase in urgency, with heart failure being one of the most severe diseases in this regard. To improve outpatient care of these patients and reduce hospitalization rates, the telemedical disease management program HerzMobil was developed in the past. OBJECTIVE: This work aims to analyze the inter-annotator variability among two professional groups (healthcare and engineering) involved in this program's annotation process of free-text clinical notes using categories. METHODS: A dataset of 1,300 text snippets was annotated by 13 annotators with different backgrounds. Inter-annotator variability and accuracy were evaluated using the F1-score and analyzed for differences between categories, annotators, and their professional backgrounds. RESULTS: The results show a significant difference between note categories concerning inter-annotator variability (p<0.0001) and accuracy (p<0.0001). However, there was no statistically significant difference between the two annotator groups, neither concerning inter-annotator variability (p=0.15) nor accuracy (p=0.84). CONCLUSION: Professional background had no significant impact on the annotation of free-text HerzMobil notes.

Classification of Clinical Notes from a Heart Failure Telehealth Network.

Heart failure is a common chronic disease which is associated with high re-hospitalization and mortality rates. Within the telemedicine-assisted transitional care disease management program HerzMobil, monitoring data such as daily measured vital parameters and various other heart failure related data are collected in a structured way. Additionally, involved healthcare professionals communicate with one another via the system using free-text clinical notes. Since manual annotation of such notes is too time-consuming for routine care applications, an automated analysis process is needed. In the present study, we established a ground truth classification of 636 randomly selected clinical notes from HerzMobil based on annotations of 9 experts with different professional background (2 physicians, 4 nurses, and 3 engineers). We analyzed the influence of the professional background on the inter annotator reliability and compared the results with the accuracy of an automated classification algorithm. We found significant differences depending on the profession and on the category. These results indicate that different professional backgrounds should be considered when selecting annotators in such scenarios.

Investigation of the time shift between wearable photoplethysmography sensors used for continuous heart rate monitoring.

In this study, we investigated the effect of time shift in heartrate measurement by wearables, which might to be used in telehealth applications for patients suffering from heart failure. Six wearables commercially available on the market were tested in a 14-hour measurement. Each wearable was tested three times by five different test persons. A reference sensor was used to test the accuracy of the wearables. We found that different types of time shifts are common in the sensors we tested: time shifts of full days, time shifts of full hours (most probably due to incorrect or unspecified time zones) and time shifts in the range of seconds to minutes (most likely stemming from averaging, data transmission, etc.). We conclude that time shifts of all manufacturers need to be corrected prior comparison of a photoplethysmography signal with other signals. However, even after correction of the time shift, the reliability of the sensors seems to be too low for application in telehealth settings. Clinical relevance- This study shows that signals from state-of-the-art wearable photoplethysmography heart rate measurements show significant time shifts and marked differences even if time shifts were corrected. This limits their utility for clinical applications.

Adherence to the Data Submission Protocol in a Diabetes Telehealth Service Pre and Post Deployment of an Adherence Optimization Module.

BACKGROUND: Telehealth services for chronic diseases are becoming more and more popular since they are expected to improve health outcomes and reduce costs. Especially for diabetes patients, life-long disease management is required. However, there are situations in a patient's life, when motivation to continue the participation in disease management programs is low and the dropout-risk is high. OBJECTIVES: We analysed if an adherence management module provided to healthcare professionals within a pre-existing diabetes telehealth service can improve the long-term adherence. METHODS: The adherence to the agreed data submission protocol was determined prior and post implementation of the adherence management module. RESULTS: Adherence to the agreed data submission protocol was higher after implementation of the adherence management module as compared to previous years. CONCLUSION: Adherence to the agreed data submission protocol can be improved by helping healthcare professionals to identify patients at risk of dropout. Further analyses are indicated to proof these results in a prospective study.

Impact Analysis of De-Identification in Clinical Notes Classification.

BACKGROUND: Clinical notes provide valuable data in telemonitoring systems for disease management. Such data must be converted into structured information to be effective in automated analysis. One way to achieve this is by classification (e.g. into categories). However, to conform with privacy regulations and concerns, text is usually de-identified. OBJECTIVES: This study investigated the effects of de-identification on classification. METHODS: Two pseudonymisation and two classification algorithms were applied to clinical messages from a telehealth system. Divergence in classification compared to clear text classification was measured. RESULTS: Overall, de-identification notably altered classification. The delicate classification algorithm was severely impacted, especially losses of sensitivity were noticeable. However, the simpler classification method was more robust and in combination with a more yielding pseudonymisation technique, had only a negligible impact on classification. CONCLUSION: The results indicate that de-identification can impact text classification and suggest, that considering de-identification during development of the classification methods could be beneficial.

Integration of Python Modules in a MATLAB-Based Predictive Analytics Toolset for Healthcare.

BACKGROUND: Python and MATLAB both are common tools used for predictive modelling applications, not only in healthcare. In our predictive modelling group, both tools are widely used. None of the two tools is optimal for all tasks along the value chain of predictive modelling in healthcare. OBJECTIVES: The aim of this study was to explore different ways to extend our MATLAB-based toolset with Python functions. METHODS: Pre-existing interfaces between MATLAB and Python have been evaluated and more comprehensive interfaces have been designed to exchange even complex data formats such as MATLAB tables. RESULTS: The interfaces have successfully been implemented and they were validated in a Natural Language Processing scenario based on free-text notes from a telehealth services for heart failure patients. CONCLUSION: Integration of Python modules in our MATLAB toolset is possible. Further improvements especially in terms of performance, are required if large datasets need to be exchanged. A big advantage of our concept is that tabular data can be exchanged between MATLAB and Python without loss and the Python functions are called dynamically via the interface.

Accuracy of Wearable Photoplethysmography Sensors for Continuous Heart Rate Monitoring in Telehealth Applications.

The demand for extended care for people suffering from heart failure is omnipresent. Wearables providing continuous heart rate measurement through optical sensors are of great interest due to their ease of use without the need for medical staff and their low cost. In this study, seven wearables were tested in fifteen measurement runs, with a duration of fourteen-hour each, and compared to a reference sensor. By calculating the Pearson correlation and the root mean square error, as well as the graphical representation by a Bland Altman plot, it was found that these wearables lack sufficient accuracy and may not be suitable for medical purposes.

Development of Algorithms for Automated Timed Up-and-Go Test Subtask and Step Frequency Analysis.

Frailty is one of the major problems associated with an aging society. Therefore, frailty assessment tools which support early detection and autonomous monitoring of the frailty status are heavily needed. One of the most used tests for functional assessment of the elderly is the "Timed Up-and-Go" test. In previous projects, we have developed an ultrasound-based device that enables performing the test autonomously. This paper described the development and validation of algorithms for detection of subtasks (stand up, walk, turn around, walk, sit down) and for step frequency estimation from the Timed Up-and-Go signals. The algorithms have been tested with an annotated test set recorded in 8 healthy subjects. The mean error for the developed subtask transition detection algorithms was in between 0.22 and 0.35 s. The mean step frequency error was 0.15 Hz. Future steps will include prospective evaluation of the algorithms with elderly people.

mHealth 6-minute walk test - accuracy for detecting clinically relevant differences in heart failure patients.

Heart failure is a serious disease which increases mortality as well as hospital admission rates for affected patients. Disease management programs supported by telehealth solutions are cost-effective approaches for reducing all-cause mortality and heart failure hospitalizations. A 6-minute walk test (6MWT) app could help heart failure patients to self-monitor their functional capacity. We have developed such an application capable of tracking the geolocation, guiding users through a 6MWT and providing the walked distance after six minutes. Besides common global navigation satellite system (GNSS) filtering methods like a Kalman filter, we have investigated the impact of positioning the device (tablet) and GNSS reception on the accuracy of the test. In a field experiment, we gathered 166 6MWT recordings with the developed mobile application. Applying the Kalman filter reduced the overall relative error from 35.5 % to 3.7 %. Wearing the tablet on the body led to significantly better results than holding it in the hand (p < .001). The average accuracy of 2.2 % of body-worn measurements was below previously defined thresholds for reliable results. It thus allows to define a procedure on how to perform and integrate an accurate 6MWT in telehealth settings for clinical decision support in heart failure patients.

Quantification of the Link between Timed Up-and-Go Test Subtasks and Contractile Muscle Properties.

Frailty and falls are a major public health problem in older adults. Muscle weakness of the lower and upper extremities are risk factors for any, as well as recurrent falls including injuries and fractures. While the Timed Up-and-Go (TUG) test is often used to identify frail members and fallers, tensiomyography (TMG) can be used as a non-invasive tool to assess the function of skeletal muscles. In a clinical study, we evaluated the correlation between the TMG parameters of the skeletal muscle contraction of 23 elderly participants (22 f, age 86.74 ± 7.88) and distance-based TUG test subtask times. TUG tests were recorded with an ultrasonic-based device. The sit-up and walking phases were significantly correlated to the contraction and delay time of the muscle vastus medialis (ρ = 0.55-0.80, p < 0.01). In addition, the delay time of the muscles vastus medialis (ρ = 0.45, p = 0.03) and gastrocnemius medialis (ρ = -0.44, p = 0.04) correlated to the sit-down phase. The maximal radial displacements of the biceps femoris showed significant correlations with the walk-forward times (ρ = -0.47, p = 0.021) and back (ρ = -0.43, p = 0.04). The association of TUG subtasks to muscle contractile parameters, therefore, could be utilized as a measure to improve the monitoring of elderly people's physical ability in general and during rehabilitation after a fall in particular. TUG test subtask measurements may be used as a proxy to monitor muscle properties in rehabilitation after long hospital stays and injuries or for fall prevention.

Natural Language Processing for Free-Text Classification in Telehealth Services: Differences Between Diabetes and Heart Failure Applications.

Telehealth services for long-term monitoring of chronically ill patients are becoming more and more common, leading to huge amounts of data collected by patients and healthcare professionals each day. While most of these data are structured, some information, especially concerning the communication between the stakeholders, is typically stored as unstructured free-texts. This paper outlines the differences in analyzing free-texts from the heart failure telehealth network HerzMobil as compared to the diabetes telehealth network DiabMemory. A total of 3,739 free-text notes from HerzMobil and 228,109 notes from DiabMemory, both written in German, were analyzed. A pre-existing, regular expression based algorithm developed for heart failure free-texts was adapted to cover also the diabetes scenario. The resulting algorithm was validated with a subset of 200 notes that were annotated by three scientists, achieving an accuracy of 92.62%. When applying the algorithm to heart failure and diabetes texts, we found various similarities but also several differences concerning the content. As a consequence, specific requirements for the algorithm were identified.

Machine Learning Based Walking Aid Detection in Timed Up-and-Go Test Recordings of Elderly Patients.

Frailty and falls are the main causes of morbidity and disability in elderly people. The Timed Up-and-Go (TUG) test has been proposed as an appropriate method for evaluating elderly individuals' risk of falling. To analyze the TUG's potential for falls prediction, we conducted a clinical study with participants aged ≥ 65 years, living in nursing homes. We harvested 138 TUG recordings with the information, if patients used a walking aid or not and developed a method to predict the use of walking aids using a Random Forest Classifier for ultrasonic based TUG test recordings. We achieved a high accuracy with an Area Under the Curve (AUC) of 96,9% using a 20% leave out evaluation strategy. Automated collection of structured data from TUG recordings - like the use of a walking aid - may help to improve fall risk tools in future.

Quantitative falls risk assessment in elderly people: results from a clinical study with distance based timed up-and-go test recordings.

OBJECTIVE: A third of people over 65 years experiences at least one fall a year. The Timed Up-and-Go (TUG) test is commonly used to assess gait and balance and to evaluate an individual's risk of falling. APPROACH: We conducted a clinical study with 46 older participants for evaluating the fall risk assessment capabilities of an ultra-sound based TUG test device. The fall protocols over a period of one year were used to classify participants as fallers and non-fallers. For frailty evaluation, state-of-the-art questionnaires were used. Fall recordings were compared to six TUG test measurements that were recorded in fallers and non-fallers. MAIN RESULTS: TUG test data were available for 39 participants (36 f, age 84.2 ± 8.2, BMI 26.0 ± 5.1). Twenty-three participants did fall at least once within the fall screening period. We fitted two different regression and probability models into a region of interest of the distance over time curve as derived from the TUG device. We found that the coefficient of determination for Gaussian bell-shaped curves (p < 0.05, AUC = 0.71) and linear regression lines (p < 0.02, AUC = 0.74) significantly separated fallers from non-fallers. Subtasks of the TUG test like the sit-up time showed near significance (p < 0.07, AUC = 0.67). SIGNIFICANCE: We found that specific features calculated from the TUG distance over time curve were significantly different between fallers and non-fallers in our study population. Automatic recording and analysis of TUG measurements could, therefore, reduce time of measurements and improve precision as compared to other methods currently being used in the assessments of fall risk.