Unobtrusive measurement of cognitive load and physiological signals in uncontrolled environments.

While individuals fail to assess their mental health subjectively in their day-to-day activities, the recent development of consumer-grade wearable devices has enormous potential to monitor daily workload objectively by acquiring physiological signals. Therefore, this work collected consumer-grade physiological signals from twenty-four participants, following a four-hour cognitive load elicitation paradigm with self-chosen tasks in uncontrolled environments and a four-hour mental workload elicitation paradigm in a controlled environment. The recorded dataset of approximately 315 hours consists of electroencephalography, acceleration, electrodermal activity, and photoplethysmogram data balanced across low and high load levels. Participants performed office-like tasks in the controlled environment (mental arithmetic, Stroop, N-Back, and Sudoku) with two defined difficulty levels and in the uncontrolled environments (mainly researching, programming, and writing emails). Each task label was provided by participants using two 5-point Likert scales of mental workload and stress and the pairwise NASA-TLX questionnaire. This data is suitable for developing real-time mental health assessment methods, conducting research on signal processing techniques for challenging environments, and developing personal cognitive load assistants.

Enhancing Preoperative Outcome Prediction: A Comparative Retrospective Case-Control Study on Machine Learning versus the International Esodata Study Group Risk Model for Predicting 90-Day Mortality in Oncologic Esophagectomy.

Risk prediction prior to oncologic esophagectomy is crucial for assisting surgeons and patients in their joint informed decision making. Recently, a new risk prediction model for 90-day mortality after esophagectomy using the International Esodata Study Group (IESG) database was proposed, allowing for the preoperative assignment of patients into different risk categories. However, given the non-linear dependencies between patient- and tumor-related risk factors contributing to cumulative surgical risk, machine learning (ML) may evolve as a novel and more integrated approach for mortality prediction. We evaluated the IESG risk model and compared its performance to ML models. Multiple classifiers were trained and validated on 552 patients from two independent centers undergoing oncologic esophagectomies. The discrimination performance of each model was assessed utilizing the area under the receiver operating characteristics curve (AUROC), the area under the precision-recall curve (AUPRC), and the Matthews correlation coefficient (MCC). The 90-day mortality rate was 5.8%. We found that IESG categorization allowed for adequate group-based risk prediction. However, ML models provided better discrimination performance, reaching superior AUROCs (0.64 [0.63-0.65] vs. 0.44 [0.32-0.56]), AUPRCs (0.25 [0.24-0.27] vs. 0.11 [0.05-0.21]), and MCCs (0.27 ([0.25-0.28] vs. 0.15 [0.03-0.27]). Conclusively, ML shows promising potential to identify patients at risk prior to surgery, surpassing conventional statistics. Still, larger datasets are needed to achieve higher discrimination performances for large-scale clinical implementation in the future.

Twenty-Four Hour Glucose Profiles and Glycemic Variability during Intermittent Religious Dry Fasting and Time-Restricted Eating in Subjects without Diabetes: A Preliminary Study.

Intermittent religious fasting increases the risk of hypo- and hyperglycemia in individuals with diabetes, but its impact on those without diabetes has been poorly investigated. The aim of this preliminary study was to examine the effects of religious Bahá'í fasting (BF) on glycemic control and variability and compare these effects with time-restricted eating (TRE). In a three-arm randomized controlled trial, 16 subjects without diabetes were assigned to a BF, TRE, or control group. Continuous glucose monitoring and food intake documentation were conducted before and during the 19 days of the intervention, and the 24 h mean glucose and glycemic variability indices were assessed. The BF and TRE groups, but not the control group, markedly reduced the daily eating window while maintaining macronutrient composition. Only the BF group decreased caloric intake (-677.8 ± 357.6 kcal, p = 0.013), body weight (-1.92 ± 0.95 kg, p = 0.011), and BMI (-0.65 ± 0.28 kg, p = 0.006). Higher maximum glucose values were observed during BF in the within-group (+1.41 ± 1.04, p = 0.039) and between-group comparisons (BF vs. control: p = 0.010; TRE vs. BF: p = 0.022). However, there were no alterations of the 24 h mean glucose, intra- and inter-day glycemic variability indices in any group. The proportions of time above and below the range (70-180 mg/dL) remained unchanged. BF and TRE do not exhibit negative effects on glycemic control and variability in subjects without diabetes.

The role of the (in)accessibility of social media data for infodemic management: a public health perspective on the situation in the European Union in March 2024.

Public health institutions rely on the access to social media data to better understand the dynamics and impact of infodemics - an overabundance of information during a disease outbreak, potentially including mis-and disinformation. The scope of the COVID-19 infodemic has led to growing concern in the public health community. The spread of harmful information or information voids may negatively impact public health. In this context, social media are of particular relevance as an integral part of our society, where much information is consumed. In this perspective paper, we discuss the current state of (in)accessibility of social media data of the main platforms in the European Union. The European Union's relatively new Digital Services Act introduces the obligation for platforms to provide data access to a wide range of researchers, likely including researchers at public health institutions without formal academic affiliation. We examined eight platforms (Facebook, Instagram, LinkedIn, Pinterest, Snapchat, TikTok, X, YouTube) affected by the new legislation in regard to data accessibility. We found that all platforms apart from TikTok offer data access through the Digital Services Act. Potentially, this presents a fundamentally new situation for research, as before the Digital Services Act, few platforms granted data access or only to very selective groups of researchers. The access regime under the Digital Services Act is, however, still evolving. Specifics such as the application procedure for researcher access are still being worked out and results can be expected in spring 2024. The impact of the Digital Services Act on research will therefore only become fully apparent in the future.

Analyzing population-level trials as N-of-1 trials: An application to gait.

Studying individual causal effects of health interventions is important whenever intervention effects are heterogeneous between study participants. Conducting N-of-1 trials, which are single-person randomized controlled trials, is the gold standard for their analysis. As an alternative method, we propose to re-analyze existing population-level studies as N-of-1 trials, and use gait as a use case for illustration. Gait data were collected from 16 young and healthy participants under fatigued and non-fatigued, as well as under single-task (only walking) and dual-task (walking while performing a cognitive task) conditions. As a reference to the N-of-1 trials approach, we first computed standard population-level ANOVA models to evaluate differences in gait parameters (stride length and stride time) across conditions. Then, we estimated the effect of the interventions on gait parameters on the individual level through Bayesian repeated-measures models, viewing each participant as their own trial, and compared the results. The results illustrated that while few overall population-level effects were visible, individual-level analyses revealed differences between participants. Baseline values of the gait parameters varied largely among all participants, and the effects of fatigue and cognitive task were also heterogeneous, with some individuals showing effects in opposite directions. These differences between population-level and individual-level analyses were more pronounced for the fatigue intervention compared to the cognitive task intervention. Following our empirical analysis, we discuss re-analyzing population studies through the lens of N-of-1 trials more generally and highlight important considerations and requirements. Our work encourages future studies to investigate individual effects using population-level data.

ARTEMIS: An alarm threshold and policy mining system for the intensive care unit.

BACKGROUND: Alarm fatigue is a major technology-induced hazard for patients and staff in intensive care units. Too many - mostly unnecessary - alarms cause desensitisation and lack of response in medical staff. Unsuitable alarm policies are one reason for alarm fatigue. But changing alarm policies is a delicate issue since it concerns patient safety. OBJECTIVE: We present ARTEMIS, a novel, computer-aided clinical decision support system for policy makers that can help to considerably improve alarm policies using data from hospital information systems. METHODS: Policy makers can use different policy components from ARTEMIS' internal library to assemble tailor-made alarm policies for their intensive care units. Alternatively, policy makers can provide even more highly customised policy components as Python functions using data the hospital information systems. This can even include machine learning models - for example for setting alarm thresholds. Finally, policy makers can evaluate their system of policies and compare the resulting alarm loads. RESULTS: ARTEMIS reports and compares numbers of alarms caused by different alarm policies for an easily adaptable target population. ARTEMIS can compare policies side-by-side and provides grid comparisons and heat maps for parameter optimisation. For example, we found that the utility of alarm delays varies based on target population. Furthermore, policy makers can introduce virtual parameters that are not in the original data by providing a formula to compute them. Virtual parameters help measuring and alarming on the right metric, even if the patient monitors do not directly measure this metric. CONCLUSION: ARTEMIS does not release the policy maker from assessing the policy from a medical standpoint. But as a knowledge discovery and clinical decision support system, it provides a strong quantitative foundation for medical decisions. At comparatively low cost of implementation, ARTEMIS can have a substantial impact on patients and staff alike - with organisational, economic, and clinical benefits for the implementing hospital.

Overcoming Data Scarcity in Human Activity Recognition.

Wearable sensors have become increasingly popular in recent years, with technological advances leading to cheaper, more widely available, and smaller devices. As a result, there has been a growing interest in applying machine learning techniques for Human Activity Recognition (HAR) in healthcare. These techniques can improve patient care and treatment by accurately detecting and analyzing various activities and behaviors. However, current approaches often require large amounts of labeled data, which can be difficult and time-consuming to obtain. In this study, we propose a new approach that uses synthetic sensor data generated by 3D engines and Generative Adversarial Networks to overcome this obstacle. We evaluate the synthetic data using several methods and compare them to real-world data, including classification results with baseline models. Our results show that synthetic data can improve the performance of deep neural networks, achieving a better F1-score for less complex activities on a known dataset by 8.4% to 73% than state-of-the-art results. However, as we showed in a self-recorded nursing activity dataset of longer duration, this effect diminishes with more complex activities. This research highlights the potential of synthetic sensor data generated from multiple sources to overcome data scarcity in HAR.

HARE: Unifying the Human Activity Recognition Engineering Workflow.

Sensor-based human activity recognition is becoming ever more prevalent. The increasing importance of distinguishing human movements, particularly in healthcare, coincides with the advent of increasingly compact sensors. A complex sequence of individual steps currently characterizes the activity recognition pipeline. It involves separate data collection, preparation, and processing steps, resulting in a heterogeneous and fragmented process. To address these challenges, we present a comprehensive framework, HARE, which seamlessly integrates all necessary steps. HARE offers synchronized data collection and labeling, integrated pose estimation for data anonymization, a multimodal classification approach, and a novel method for determining optimal sensor placement to enhance classification results. Additionally, our framework incorporates real-time activity recognition with on-device model adaptation capabilities. To validate the effectiveness of our framework, we conducted extensive evaluations using diverse datasets, including our own collected dataset focusing on nursing activities. Our results show that HARE's multimodal and on-device trained model outperforms conventional single-modal and offline variants. Furthermore, our vision-based approach for optimal sensor placement yields comparable results to the trained model. Our work advances the field of sensor-based human activity recognition by introducing a comprehensive framework that streamlines data collection and classification while offering a novel method for determining optimal sensor placement.

Optimization of data pre-processing methods for time-series classification of electroencephalography data.

The performance of time-series classification of electroencephalographic data varies strongly across experimental paradigms and study participants. Reasons are task-dependent differences in neuronal processing and seemingly random variations between subjects, amongst others. The effect of data pre-processing techniques to ameliorate these challenges is relatively little studied. Here, the influence of spatial filter optimization methods and non-linear data transformation on time-series classification performance is analyzed by the example of high-frequency somatosensory evoked responses. This is a model paradigm for the analysis of high-frequency electroencephalography data at a very low signal-to-noise ratio, which emphasizes the differences of the explored methods. For the utilized data, it was found that the individual signal-to-noise ratio explained up to 74% of the performance differences between subjects. While data pre-processing was shown to increase average time-series classification performance, it could not fully compensate the signal-to-noise ratio differences between the subjects. This study proposes an algorithm to prototype and benchmark pre-processing pipelines for a paradigm and data set at hand. Extreme learning machines, Random Forest, and Logistic Regression can be used quickly to compare a set of potentially suitable pipelines. For subsequent classification, however, machine learning models were shown to provide better accuracy.

State Space Modeling of Event Count Time Series.

This paper proposes a class of algorithms for analyzing event count time series, based on state space modeling and Kalman filtering. While the dynamics of the state space model is kept Gaussian and linear, a nonlinear observation function is chosen. In order to estimate the states, an iterated extended Kalman filter is employed. Positive definiteness of covariance matrices is preserved by a square-root filtering approach, based on singular value decomposition. Non-negativity of the count data is ensured, either by an exponential observation function, or by a newly introduced "affinely distorted hyperbolic" observation function. The resulting algorithm is applied to time series of the daily number of seizures of drug-resistant epilepsy patients. This number may depend on dosages of simultaneously administered anti-epileptic drugs, their superposition effects, delay effects, and unknown factors, making the objective analysis of seizure counts time series arduous. For the purpose of validation, a simulation study is performed. The results of the time series analysis by state space modeling, using the dosages of the anti-epileptic drugs as external control inputs, provide a decision on the effect of the drugs in a particular patient, with respect to reducing or increasing the number of seizures.

SONAR, a nursing activity dataset with inertial sensors.

Accurate and comprehensive nursing documentation is essential to ensure quality patient care. To streamline this process, we present SONAR, a publicly available dataset of nursing activities recorded using inertial sensors in a nursing home. The dataset includes 14 sensor streams, such as acceleration and angular velocity, and 23 activities recorded by 14 caregivers using five sensors for 61.7 hours. The caregivers wore the sensors as they performed their daily tasks, allowing for continuous monitoring of their activities. We additionally provide machine learning models that recognize the nursing activities given the sensor data. In particular, we present benchmarks for three deep learning model architectures and evaluate their performance using different metrics and sensor locations. Our dataset, which can be used for research on sensor-based human activity recognition in real-world settings, has the potential to improve nursing care by providing valuable insights that can identify areas for improvement, facilitate accurate documentation, and tailor care to specific patient conditions.

DUO-GAIT: A gait dataset for walking under dual-task and fatigue conditions with inertial measurement units.

In recent years, there has been a growing interest in developing and evaluating gait analysis algorithms based on inertial measurement unit (IMU) data, which has important implications, including sports, assessment of diseases, and rehabilitation. Multi-tasking and physical fatigue are two relevant aspects of daily life gait monitoring, but there is a lack of publicly available datasets to support the development and testing of methods using a mobile IMU setup. We present a dataset consisting of 6-minute walks under single- (only walking) and dual-task (walking while performing a cognitive task) conditions in unfatigued and fatigued states from sixteen healthy adults. Especially, nine IMUs were placed on the head, chest, lower back, wrists, legs, and feet to record under each of the above-mentioned conditions. The dataset also includes a rich set of spatio-temporal gait parameters that capture the aspects of pace, symmetry, and variability, as well as additional study-related information to support further analysis. This dataset can serve as a foundation for future research on gait monitoring in free-living environments.

Comparison of Bayesian Networks, G-estimation and linear models to estimate causal treatment effects in aggregated N-of-1 trials with carry-over effects.

BACKGROUND: The aggregation of a series of N-of-1 trials presents an innovative and efficient study design, as an alternative to traditional randomized clinical trials. Challenges for the statistical analysis arise when there is carry-over or complex dependencies of the treatment effect of interest. METHODS: In this study, we evaluate and compare methods for the analysis of aggregated N-of-1 trials in different scenarios with carry-over and complex dependencies of treatment effects on covariates. For this, we simulate data of a series of N-of-1 trials for Chronic Nonspecific Low Back Pain based on assumed causal relationships parameterized by directed acyclic graphs. In addition to existing statistical methods such as regression models, Bayesian Networks, and G-estimation, we introduce a carry-over adjusted parametric model (COAPM). RESULTS: The results show that all evaluated existing models have a good performance when there is no carry-over and no treatment dependence. When there is carry-over, COAPM yields unbiased and more efficient estimates while all other methods show some bias in the estimation. When there is known treatment dependence, all approaches that are capable to model it yield unbiased estimates. Finally, the efficiency of all methods decreases slightly when there are missing values, and the bias in the estimates can also increase. CONCLUSIONS: This study presents a systematic evaluation of existing and novel approaches for the statistical analysis of a series of N-of-1 trials. We derive practical recommendations which methods may be best in which scenarios.

Food Choices after Cognitive Load: An Affective Computing Approach.

Psychology and nutritional science research has highlighted the impact of negative emotions and cognitive load on calorie consumption behaviour using subjective questionnaires. Isolated studies in other domains objectively assess cognitive load without considering its effects on eating behaviour. This study aims to explore the potential for developing an integrated eating behaviour assistant system that incorporates cognitive load factors. Two experimental sessions were conducted using custom-developed experimentation software to induce different stimuli. During these sessions, we collected 30 h of physiological, food consumption, and affective states questionnaires data to automatically detect cognitive load and analyse its effect on food choice. Utilising grid search optimisation and leave-one-subject-out cross-validation, a support vector machine model achieved a mean classification accuracy of 85.12% for the two cognitive load tasks using eight relevant features. Statistical analysis was performed on calorie consumption and questionnaire data. Furthermore, 75% of the subjects with higher negative affect significantly increased consumption of specific foods after high-cognitive-load tasks. These findings offer insights into the intricate relationship between cognitive load, affective states, and food choice, paving the way for an eating behaviour assistant system to manage food choices during cognitive load. Future research should enhance system capabilities and explore real-world applications.

Online Learning for Wearable EEG-Based Emotion Classification.

Giving emotional intelligence to machines can facilitate the early detection and prediction of mental diseases and symptoms. Electroencephalography (EEG)-based emotion recognition is widely applied because it measures electrical correlates directly from the brain rather than indirect measurement of other physiological responses initiated by the brain. Therefore, we used non-invasive and portable EEG sensors to develop a real-time emotion classification pipeline. The pipeline trains different binary classifiers for Valence and Arousal dimensions from an incoming EEG data stream achieving a 23.9% (Arousal) and 25.8% (Valence) higher F1-Score on the state-of-art AMIGOS dataset than previous work. Afterward, the pipeline was applied to the curated dataset from 15 participants using two consumer-grade EEG devices while watching 16 short emotional videos in a controlled environment. Mean F1-Scores of 87% (Arousal) and 82% (Valence) were achieved for an immediate label setting. Additionally, the pipeline proved to be fast enough to achieve predictions in real-time in a live scenario with delayed labels while continuously being updated. The significant discrepancy from the readily available labels on the classification scores leads to future work to include more data. Thereafter, the pipeline is ready to be used for real-time applications of emotion classification.

A Standardized Clinical Data Harmonization Pipeline for Scalable AI Application Deployment (FHIR-DHP): Validation and Usability Study.

BACKGROUND: Increasing digitalization in the medical domain gives rise to large amounts of health care data, which has the potential to expand clinical knowledge and transform patient care if leveraged through artificial intelligence (AI). Yet, big data and AI oftentimes cannot unlock their full potential at scale, owing to nonstandardized data formats, lack of technical and semantic data interoperability, and limited cooperation between stakeholders in the health care system. Despite the existence of standardized data formats for the medical domain, such as Fast Healthcare Interoperability Resources (FHIR), their prevalence and usability for AI remain limited. OBJECTIVE: In this paper, we developed a data harmonization pipeline (DHP) for clinical data sets relying on the common FHIR data standard. METHODS: We validated the performance and usability of our FHIR-DHP with data from the Medical Information Mart for Intensive Care IV database. RESULTS: We present the FHIR-DHP workflow in respect of the transformation of "raw" hospital records into a harmonized, AI-friendly data representation. The pipeline consists of the following 5 key preprocessing steps: querying of data from hospital database, FHIR mapping, syntactic validation, transfer of harmonized data into the patient-model database, and export of data in an AI-friendly format for further medical applications. A detailed example of FHIR-DHP execution was presented for clinical diagnoses records. CONCLUSIONS: Our approach enables the scalable and needs-driven data modeling of large and heterogenous clinical data sets. The FHIR-DHP is a pivotal step toward increasing cooperation, interoperability, and quality of patient care in the clinical routine and for medical research.

Protocol for a Randomized Crossover Trial to Evaluate the Effect of Soft Brace and Rigid Orthosis on Performance and Readiness to Return to Sport Six Months Post-ACL-Reconstruction.

A randomized crossover trial was designed to investigate the influence of muscle activation and strength on functional stability/control of the knee joint, to determine whether bilateral imbalances still occur six months after successful anterior cruciate ligament reconstruction (ACLR), and to analyze whether the use of orthotic devices changes the activity onset of these muscles. Furthermore, conclusions on the feedforward and feedback mechanisms are highlighted. Therefore, twenty-eight patients will take part in a modified Back in Action (BIA) test battery at an average of six months after a primary unilateral ACLR, which used an autologous ipsilateral semitendinosus tendon graft. This includes double-leg and single-leg stability tests, double-leg and single-leg countermovement jumps, double-leg and single-leg drop jumps, a speedy jump test, and a quick feet test. During the tests, gluteus medius and semitendinosus muscle activity are analyzed using surface electromyography (sEMG). Motion analysis is conducted using Microsoft Azure DK and 3D force plates. The tests are performed while wearing knee rigid orthosis, soft brace, and with no aid, in random order. Additionally, the range of hip and knee motion and hip abductor muscle strength under isometric conditions are measured. Furthermore, patient-rated outcomes will be assessed.

Computational approaches to alleviate alarm fatigue in intensive care medicine: A systematic literature review.

Patient monitoring technology has been used to guide therapy and alert staff when a vital sign leaves a predefined range in the intensive care unit (ICU) for decades. However, large amounts of technically false or clinically irrelevant alarms provoke alarm fatigue in staff leading to desensitisation towards critical alarms. With this systematic review, we are following the Preferred Reporting Items for Systematic Reviews (PRISMA) checklist in order to summarise scientific efforts that aimed to develop IT systems to reduce alarm fatigue in ICUs. 69 peer-reviewed publications were included. The majority of publications targeted the avoidance of technically false alarms, while the remainder focused on prediction of patient deterioration or alarm presentation. The investigated alarm types were mostly associated with heart rate or arrhythmia, followed by arterial blood pressure, oxygen saturation, and respiratory rate. Most publications focused on the development of software solutions, some on wearables, smartphones, or headmounted displays for delivering alarms to staff. The most commonly used statistical models were tree-based. In conclusion, we found strong evidence that alarm fatigue can be alleviated by IT-based solutions. However, future efforts should focus more on the avoidance of clinically non-actionable alarms which could be accelerated by improving the data availability. Systematic Review Registration: https://www.crd.york.ac.uk/prospero/display_record.php?ID=CRD42021233461, identifier: CRD42021233461.

Wearable electroencephalography and multi-modal mental state classification: A systematic literature review.

BACKGROUND: Wearable multi-modal time-series classification applications outperform their best uni-modal counterparts and hold great promise. A modality that directly measures electrical correlates from the brain is electroencephalography. Due to varying noise sources, different key brain regions, key frequency bands, and signal characteristics like non-stationarity, techniques for data pre-processing and classification algorithms are task-dependent. METHOD: Here, a systematic literature review on mental state classification for wearable electroencephalography is presented. Four search terms in different combinations were used for an in-title search. The search was executed on the 29th of June 2022, across Google Scholar, PubMed, IEEEXplore, and ScienceDirect. 76 most relevant publications were set into context as the current state-of-the-art in mental state time-series classification. RESULTS: Pre-processing techniques, features, and time-series classification models were analyzed. Across publications, a window length of one second was mainly chosen for classification and spectral features were utilized the most. The achieved performance per time-series classification model is analyzed, finding linear discriminant analysis, decision trees, and k-nearest neighbors models outperform support-vector machines by a factor of up to 1.5. A historical analysis depicts future trends while under-reported aspects relevant to practical applications are discussed. CONCLUSIONS: Five main conclusions are given, covering utilization of available area for electrode placement on the head, most often or scarcely utilized features and time-series classification model architectures, baseline reporting practices, as well as explainability and interpretability of Deep Learning. The importance of a 'test battery' assessing the influence of data pre-processing and multi-modality on time-series classification performance is emphasized.

Defending against Reconstruction Attacks through Differentially Private Federated Learning for Classification of Heterogeneous Chest X-ray Data.

Privacy regulations and the physical distribution of heterogeneous data are often primary concerns for the development of deep learning models in a medical context. This paper evaluates the feasibility of differentially private federated learning for chest X-ray classification as a defense against data privacy attacks. To the best of our knowledge, we are the first to directly compare the impact of differentially private training on two different neural network architectures, DenseNet121 and ResNet50. Extending the federated learning environments previously analyzed in terms of privacy, we simulated a heterogeneous and imbalanced federated setting by distributing images from the public CheXpert and Mendeley chest X-ray datasets unevenly among 36 clients. Both non-private baseline models achieved an area under the receiver operating characteristic curve (AUC) of 0.94 on the binary classification task of detecting the presence of a medical finding. We demonstrate that both model architectures are vulnerable to privacy violation by applying image reconstruction attacks to local model updates from individual clients. The attack was particularly successful during later training stages. To mitigate the risk of a privacy breach, we integrated Rényi differential privacy with a Gaussian noise mechanism into local model training. We evaluate model performance and attack vulnerability for privacy budgets ε∈{1,3,6,10}. The DenseNet121 achieved the best utility-privacy trade-off with an AUC of 0.94 for ε=6. Model performance deteriorated slightly for individual clients compared to the non-private baseline. The ResNet50 only reached an AUC of 0.76 in the same privacy setting. Its performance was inferior to that of the DenseNet121 for all considered privacy constraints, suggesting that the DenseNet121 architecture is more robust to differentially private training.