Self-testing has been proven to significantly improve not only simple learning outcomes, but also higher-order skills such as clinical reasoning in medical students. Previous studies have shown that self-testing was especially beneficial when it was presented with feedback, which leaves the question whether an immediate and personalized feedback further encourages this effect. Therefore, we hypothesised that individual feedback has a greater effect on learning outcomes, compared to generic feedback. In a randomised cross-over trial, German medical students were invited to voluntarily answer daily key-feature questions via an App. For half of the items they received a generalised feedback by an expert, while the feedback on the other half was generated immediately through ChatGPT. After the intervention, the students participated in a mandatory exit exam. Those participants who used the app more frequently experienced a better learning outcome compared to those who did not use it frequently, even though this finding was only examined in a correlative nature. The individual ChatGPT generated feedback did not show a greater effect on exit exam scores compared to the expert comment (51.8 ± 22.0% vs. 55.8 ± 22.8%; p = 0.06). This study proves the concept of providing personalised feedback on medical questions. Despite the promising results, improved prompting and further development of the application seems necessary to strengthen the possible impact of the personalised feedback. Our study closes a research gap and holds great potential for further use not only in medicine but also in other academic fields.

There is a growing consensus in the global health community that the use of communication technologies will be an essential factor in ensuring universal health coverage of the world’s population. New technologies can only be used profitably if their accuracy is sufficient. Therefore, we explore the feasibility of using Apple’s ARKit technology to accurately measure the distance from the user’s eye to their smartphone screen. We developed an iOS application for measuring eyes-to-phone distances in various angles, using the built-in front-facing-camera and TrueDepth sensor. The actual position of the phone is precisely controlled and recorded, by fixing the head position and placing the phone in a robotic arm. Our results indicate that ARKit is capable of producing accurate measurements, with overall errors ranging between 0.88% and 9.07% from the actual distance, across various head positions. The accuracy of ARKit may be impacted by several factors such as head size, position, device model, and temperature. Our findings suggest that ARKit is a useful tool in the development of applications aimed at preventing eye damage caused by smartphone use.

Background: Identifying high-risk individuals is crucial for preventing cardiovascular diseases (CVDs). Currently, risk assessment is mostly performed by physicians. Mobile health apps could help decouple the determination of risk from medical resources by allowing unrestricted self-assessment. The respective test results need to be interpretable for laypersons. Objective: Together with a patient organization, we aimed to design a digital risk calculator that allows people to individually assess and optimize their CVD risk. The risk calculator was integrated into the mobile health app HerzFit, which provides the respective background information. Methods: To cover a broad spectrum of individuals for both primary and secondary prevention, we integrated the respective scores (Framingham 10-year CVD, Systematic Coronary Risk Evaluation 2, Systematic Coronary Risk Evaluation 2 in Older Persons, and Secondary Manifestations Of Arterial Disease) into a single risk calculator that was recalibrated for the German population. In primary prevention, an individual’s heart age is estimated, which gives the user an easy-to-understand metric for assessing cardiac health. For secondary prevention, the risk of recurrence was assessed. In addition, a comparison of expected to mean and optimal risk levels was determined. The risk calculator is available free of charge. Data safety is ensured by processing the data locally on the users’ smartphones. Results: Offering a risk calculator to the general population requires the use of multiple instruments, as each provides only a limited spectrum in terms of age and risk distribution. The integration of 4 internationally recommended scores allows risk calculation in individuals aged 30 to 90 years with and without CVD. Such integration requires recalibration and harmonization to provide consistent and plausible estimates. In the first 14 months after the launch, the HerzFit calculator was downloaded more than 96,000 times, indicating great demand. Public information campaigns proved effective in publicizing the risk calculator and contributed significantly to download numbers. Conclusions: The HerzFit calculator provides CVD risk assessment for the general population. The public demonstrated great demand for such a risk calculator as it was downloaded up to 10,000 times per month, depending on campaigns creating awareness for the instrument.

In human pose estimation, a comprehensive evaluation of state-of-the-art frameworks is necessary to advance both research and practical applications. This paper presents a thorough review of state-of-the-art 2D and 3D human pose estimation frameworks, analyzing 118 papers and four GitHub repositories, with a focus on frameworks made since 2019. The following frameworks are chosen based on predefined inclusion criteria: AlphaPose, Detectron2, MediaPipe, MeTRAbs, MHFormer, MMPose, MoveNet, OpenPifPaf, OpenPifPaf-vita, OpenPose, PoseFormerV2, rtmlib, StridedTransformer-Pose3D, ultralytics (YOLOv8), ViTPose, and YOLOv7. This paper evaluates these 16 frameworks on an existing, unpublished dataset consisting of exercise videos recorded with a monocular RGB camera and synchronized gold-standard motion capture data. The dataset includes videos of nine individuals performing eight exercises, recorded from two camera views with different planar angles. The analysis evaluates joint angle performance of the frameworks using weighted mean absolute error and weighted intraclass correlation coefficient as quantitative metrics. MeTRAbs emerged as the best overall framework, while AlphaPose, rtmlib, and YOLOv7 were the top 2D performers. The used code is available open source.¹.

Background: Digital short cognitive tests administered by medical assistants (MAs) in general practitioners’ (GPs) practices have great potential for the timely identification of patients with dementia, because they can lead to targeted specialist referrals or to immediate reassurance of patients regarding their perceived concerns. However, integration of this testing approach into clinical practice requires good usability for the test itself, especially for cognitively impaired older adults. Objective: In this implementation study, the digital version of the Montreal Cognitive Assessment (MoCA) Duo was conducted by MAs in general practice. We tested if the interaction with the test is associated with usability problems for the patients and aimed to find additional relevant constructs that should be considered for the potential implementation of such digital tests into clinical practice. We focused the study on subjective success, usability, and workload as well as their association with the result of the cognitive test to assess whether the MoCA Duo can be implemented into general practice. Methods: In total, 10 GPs took part in the study. Within their practices, 299 GP patients (aged 51-97 years) with cognitive concerns completed the MoCA Duo administered by MAs. Subsequently, patients and MAs completed digital questionnaires regarding the interaction with the app. Usability was measured using the adapted System Usability Scale, and perceived workload using the National Aeronautics and Space Administration Task Load Index. For the perceived workload, we included an assessment of the patient by the MA. Results of the MoCA Duo were supplied to the GPs for their consultation with the patient. Results: The results indicated good usability for the MoCA Duo. Self-assessment by the patients indicated that 64% (191/299) could perform in the test to the best of their ability, affected by their MoCA score. We found significant higher usability ratings by patients with better MoCA scores as well as by younger patients. Furthermore, the perceived workload showed overall medium workload. We found significant correlations between the subjective perceived workload of the patients and the assessment by MAs. Self-assessments as well as assessments by the MAs were significantly influenced by the MoCA scores and the age of the participants. Conclusions: The results indicate good usability of the digital MoCA within the sample, supporting the idea that the resulting scores are adequate to assess cognitive status without dependence on technological affinity. Furthermore, the results highlight the relevance of heterogenous samples for comparable evaluation studies, based on the significant effect of cognitive status and age on usability and workload.

Facial features hold information about a person’s emotions, motor function, or genetic defects. Since most current mobile devices are capable of real-time face detection using cameras and depth sensors, real-time facial analysis can be utilized in several mobile use cases. Understanding the real-time emotion recognition capabilities of device sensors and frameworks is vital for developing new, valid applications. Therefore, we evaluated on-device emotion recognition using Apple’s ARKit on an iPhone 14 Pro. A native app elicited 36 blend shape-specific movements and 7 discrete emotions from $N=31$ healthy adults. Per frame, standardized ARKit blend shapes were classified using a prototype-based cosine similarity metric; performance was summarized as accuracy and area under the receiver operating characteristic curves. Cosine similarity achieved an overall accuracy of $68.3\%$, exceeding the mean of three human raters ($58.9\%$; $+9.4$ percentage points, ≈16% relative). Per-emotion accuracy was highest for joy, fear, sadness, and surprise, and competitive for anger, disgust, and contempt. AUCs were ≥0.84 for all classes. The method runs in real time on-device using only vector operations, preserving privacy and minimizing compute. These results indicate that a simple, interpretable cosine-similarity classifier over ARKit blend shapes delivers human-comparable, real-time facial emotion recognition on commodity hardware, supporting privacy-preserving mobile applications.