We present a pioneering investigation into the application of deep learning techniques to analyze histopathological images for addressing the substantial challenge of automated prognostic prediction. Prognostic prediction poses a unique challenge as the ground truth labels are inherently weak, and the model must anticipate future events that are not directly observable in the image. To address this challenge, we propose a novel three-part framework comprising of a convolutional network based tissue segmentation algorithm for region of interest delineation, a contrastive learning module for feature extraction, and a nested multiple instance learning classification module. Our study explores the significance of various regions of interest within the histopathological slides and exploits diverse learning methods in real-world clinical scenarios. The pipeline is initially validated on artificially generated data and a simpler diagnostic task. Transitioning to prognostic prediction, tasks become more challenging. Employing bladder cancer as use case, our best models yield an AUC of 0.721 and 0.678 for recurrence and treatment outcome prediction respectively for a private data cohort. Altogether, this research serves as an initial investigation on the shortcomings of histopathological image analysis for treatment outcome prediction.

Alzheimer's disease (AD) is a neurodegenerative condition and the most common form of dementia. Recent developments in AD treatment call for robust diagnostic tools to facilitate medical decision-making. Despite progress for early diagnostic tests, there remains uncertainty about clinical use. Structural magnetic resonance imaging (MRI), as a readily available imaging tool in the current AD diagnostic pathway, in combination with artificial intelligence, offers opportunities of added value beyond symptomatic evaluation. However, MRI studies in AD tend to suffer from small datasets and consequently limited generalizability. Although ensemble models take advantage of the strengths of several models to improve performance and generalizability, there is little knowledge of how the different ensemble models compare performance-wise and the relationship between detection performance and model calibration. The latter is especially relevant for clinical translatability. In our study, we applied three ensemble decision strategies with three different deep learning architectures for multi-class AD detection with structural MRI. For two of the three architectures, the weighted average was the best decision strategy in terms of balanced accuracy and calibration error. In contrast to the base models, the results of the ensemble models showed that the best detection performance corresponded to the lowest calibration error, independent of the architecture. For each architecture, the best ensemble model reduced the estimated calibration error compared to the base model average from (1) 0.174±0.01 to 0.164±0.04, (2) 0.182±0.02 to 0.141±0.04, and (3) 0.269±0.08 to 0.240±0.04 and increased the balanced accuracy from (1) 0.527±0.05 to 0.608±0.06, (2) 0.417±0.03 to 0.456±0.04, and (3) 0.348±0.02 to 0.371±0.03.

Background Late Gadolinium-enhancement in cardiac magnetic resonance imaging (LGE-CMR) is the gold standard for assessing myocardial infarction (MI) size. Texture-based probability mapping (TPM) is a novel machine learning-based analysis of LGE images of myocardial injury. The ability of TPM to assess acute myocardial injury has not been determined. Objectives This proof-of-concept study aimed to determine how TPM responds to the dynamic changes in myocardial injury during one-year follow-up after a first-time revascularized acute MI. Methods 41 patients with first-time acute ST-elevation MI and single-vessel occlusion underwent successful PCI. LGE-CMR images were obtained 2 days, 1 week, 2 months, and 1 year following MI. TPM size was compared with manual LGE-CMR based MI size, LV remodeling, and biomarkers. Results TPM size remained larger than MI by LGE-CMR at all time points, decreasing from 2 days to 2 months (p < 0.001) but increasing from 2 months to 1 year (p < 0.01). TPM correlated strongly with peak Troponin T (p < 0.001) and NT-proBNP (p < 0.001). At 1 week, 2 months, and 1 year, TPM showed a stronger correlation with NT-proBNP than MI size by LGE-CMR. Analyzing all collected pixels from 2 months to 1 year revealed a general increase in pixel scar probability in both the infarcted and non-infarcted regions. Conclusion This proof-of-concept study suggests that TPM may offer additional insights into myocardial alterations in both infarcted and non-infarcted regions following acute MI. These findings indicate a potential role for TPM in assessing the overall myocardial response to infarction and the subsequent healing and remodeling process.

BACKGROUND:Patients who achieve return of spontaneous circulation (ROSC) after in-hospital cardiac arrest (IHCA) may re-arrest. This phenomenon has not been sufficiently investigated. The aim of this study was to examine the immediate (1-min) and short-term (20-min) risks of re-arrest in IHCA. METHODS:We retrospectively analyzed four datasets of IHCA episodes, comprising defibrillator recordings collected between 2002 and 2022. Re-arrest was defined as the resumption of chest compressions following a period of ROSC after cardiac arrest of any duration. Parametric models were applied to calculate the immediate risk of re-arrest. In addition, we estimated the short-term risk of re-arrest within 20 min. RESULTS:In 763 episodes of IHCA, we observed 316 re-arrests: 68% to pulseless electrical activity (PEA), 25% to ventricular fibrillation/ventricular tachycardia (VF/VT), and 7% to asystole. Most re-arrests occurred with the same rhythm as in the initial arrest. When ROSC was achieved from a non-shockable rhythm, the risk of re-arrest to a non-shockable rhythm was initially 2% per minute and decreased to 1% per minute after 9 min. The corresponding risk of re-arrest to VF/VT was constant at 2% per minute. If ROSC was obtained from a shockable rhythm, the risk of re-arrest to a shockable rhythm was initially 5% per minute, decreasing to 4% per minute after 9 min. The corresponding risk to a non-shockable rhythm was constant at 1% per minute. The risk of re-arrest within 20 min was 27%, and the overall risk of at least one re-arrest per episode was 33%. CONCLUSIONS:The immediate risk of re-arrest was approximately 2% per minute, with the highest risk occurring as a reversion to VF/VT if ROSC was obtained from VF/VT. The risk of re-arrest within 20 min of the initial arrest was 27%, and the overall risk of at least one re-arrest per episode was 33%.

Cognitive impairments in attention are prevalent among patients with Alzheimer’s disease (AD), Parkinson’s disease (PD), dementia with Lewy bodies (DLB), and Parkinson’s disease dementia (PDD). Electroencephalograms (EEGs), particularly event-related potentials (ERPs), provide valuable insights into these impairments. This study explores the use of Hjorth descriptors—activity, mobility, and complexity—from ERP signals to automatically classify these cognitive disorders. We analyzed EEG data from five subject groups (DLB, PD, PDD, AD, and healthy controls) using k-nearest neighbors, random forest, and gradient boosting classifiers. Data were collected from 90 subjects and each participant underwent neuropsychological assessments and EEG recordings during auditory oddball-distractor tasks. ERP segments were made based on three different events used during the paradigm. Our findings revealed significant differences in the Hjorth descriptors when compared with healthy controls, particularly in the PDD group, where decreased mobility indicated lower mental activity or alertness. The Random Forest classifier outperformed other methods, emphasizing its potential for effective differentiation of cognitive disorders. This study highlights the utility of EEG and machine learning in the early detection and classification of neurodegenerative diseases, offering valuable insights for better patient management.

Coronary artery calcification (CAC) due to coronary artery disease (CAD) poses significant risks of heart attack, sudden cardiac death, and other cardiac complications. CAC reflects progressive CAD, which in sonic individuals may be accelerated by high-intensity exercise. Identifying individuals with CAC and determining safe levels of physical exercise is therefore important. The most common way to identify CAC is by coronary computed tomography angiography (CCTA), but due to limited capacity for CCTA assessment, it is highly valuable to develop tools to proactively identify individuals that may benefit from this assessment. Previous studies have demonstrated significant differences in the physiological response to exercise between individuals with and without CAC. In the present study, we applied machine learning methods to physiological data acquired in relation to prolonged high-intensity exercise in individuals without symptoms or signs of CAD. All individuals were assessed by CCTA after exercise. Various dimensionality reduction methods and classification algorithms were assessed, applying nested cross validation for hyperparameter optimization and model testing. The best performing model predicted the presence of CAC with an accuracy of 84%, correctly identifying 86% of the individuals with CAC. Feature subset selection was also carried out to determine the most important input parameters, highlighting the most important physiological parameters as age and blood pressure measured directly after high-intensity exercise. The present findings support the use of machine learning methods on physiological measurements and sensor data to identify individuals who may benefit from CCTA assessment. The best-performing model showed strong predictive power of presence of CAC using only age, blood pressure, body mass index and heart rate variability as input features.

Deprivation of oxygen in an infant during and after birth leads to birth asphyxia, which is considered one of the leading causes of death in the neonatal period. Adequate resuscitation activities are performed immediately after birth to save the majority of newborns. The primary resuscitation activities include ventilation, stimulation, drying, suction, and chest compression. While resuscitation guidelines exist, little research has been conducted on measured resuscitation episodes. Objective data collected for measuring and registration of the executed resuscitation activities can be used to generate temporal timelines. This paper is primarily aimed to introduce methods for analyzing newborn resuscitation activity timelines, through visualization, aggregation, redundancy and dimensionality reduction. We are using two datasets: 1) from Stavanger University Hospital with 108 resuscitation episodes, and 2) from Haydom Lutheran Hospital with 76 episodes. The resuscitation activity timelines were manually annotated, but in future work we will use the proposed method on automatically generated timelines from video and sensor data. We propose an encoding generator with unique codes for combination of activities. A visualization of aggregated episodes is proposed using sparse nearest neighbor graph, shown to be useful to compare datasets and give insights. Finally, we propose a method consisting of an autoencoder trained for reducing redundancy in encoded resuscitation timeline descriptions, followed by a neighborhood component analysis for dimensionality reduction. Visualization of the resulting features shows very good class separability and potential for clustering the resuscitation files according to the outcome of the newborns as dead, admitted to NICU or normal. This shows great potential for extracting important resuscitation patterns when tested on larger datasets.

Endurance exercise is associated with increased life duration and improved life quality. Paradoxically, high exercise intensity is also associated with increased coronary artery calcification (CAC) and a small but significant increased risk of adverse cardiac events during exercise. The mechanisms underlying the development of CAC during prolonged high-intensity endurance exercise are unknown. This study aims to determine if there are differences in cardiovascular haemodynamic measures and heart rate variability (HRV) in individuals with (CAC+) and without CAC (CAC-). Hemodynamic measures from 56 healthy, middle-aged (median [interquartile range] 51 [43-58] years) individuals (41 men/15 women) participating in a 91 km [251.2 [217.2-271.6] min] leisure sport mountain bike race were included in this study. Twenty-five participants (20 men/5 women) were classified as CAC+ based on coronary computed tomographic assessment. Haemodynamic measures and HRV were quantified at the top of the hardest hill (THH) during the last quarter of the race. At the top of THH, CAC+ individuals had significantly higher systolic blood pressure (SBP) (235 [225-245] mmHg vs. 220 [193-238] mmHg, P = 0.008), higher diastolic blood pressure (DBP) (105 [95-110] mmHg vs. 95 [85-110] mmHg, P = 0.006), higher pulse pressure (130 [125-140] mmHg vs. 123 [110-130] mmHg, P = 0.039), higher mean rate pressure product (33,882 [30,872-35,053] bpm × mmHg vs. 31,028 [27,392-33,047] bpm × mmHg, P = 0.028), and larger increase in DBP from baseline (20 [20-30] mmHg vs. 10 [0-20] mmHg, P = 0.001), compared with CAC- individuals. Further, CAC+ participants showed a significant reduction in the low-frequency component of HRV (HRVLF) (6.3 [2.4-11.5] ms2 vs. 12.4 [6.8-20.2] ms2, P = 0.044). In multivariable analysis, HRVLF was an independent predictor of the presence of CAC even after adjusting for established risk factors of atherosclerosis: age, sex, body mass index, maximum heart rate, V ̇ O 2 max ${{\dot{V}}_{{{{\mathrm{O}}}_{\mathrm{2}}}{\mathrm{max}}}}$ , smoking, resting SBP and resting DBP. CAC+ individuals had significant alterations in haemodynamic measures and HRVLF following prolonged high-intensity endurance exercise compared with individuals without CAC. HRVLF was an independent predictor of CAC, suggesting an adverse autonomic response to high-intensity endurance exercise in individuals with CAC.

BackgroundA defibrillator should be connected to all patients receiving cardiopulmonary resuscitation (CPR) to allow early defibrillation. The defibrillator will collect signal data such as the electrocardiogram (ECG), thoracic impedance and end-tidal CO2, which allows for research on how patients demonstrate different responses to CPR. The aim of this review is to give an overview of methodological challenges and opportunities in using defibrillator data for research.MethodsThe successful collection of defibrillator files has several challenges. There is no scientific standard on how to store such data, which have resulted in several proprietary industrial solutions. The data needs to be exported to a software environment where signal filtering and classifications of ECG rhythms can be performed. This may be automated using different algorithms and artificial intelligence (AI). The patient can be classified being in ventricular fibrillation or -tachycardia, asystole, pulseless electrical activity or having obtained return of spontaneous circulation. How this dynamic response is time-dependent and related to covariates can be handled in several ways. These include Aalen’s linear model, Weibull regression and joint models.ConclusionsThe vast amount of signal data from defibrillator represents promising opportunities for the use of AI and statistical analysis to assess patient response to CPR. This may provide an epidemiologic basis to improve resuscitation guidelines and give more individualized care. We suggest that an international working party is initiated to facilitate a discussion on how open formats for defibrillator data can be accomplished, that obligates industrial partners to further develop their current technological solutions.

Background Physical exercise is associated with increased life duration and improved life quality. Paradoxically, high exercise intensity is also associated with increased coronary artery calcification (CAC) and a small but significant increased risk of myocardial infarction and cardiac death during high-intensity exercise. The mechanisms and the clinical implications of the association between prolonged high exercise intensity and increased CAC are unknown. Aim This study aims to determine if there are differences in cardiovascular hemodynamic measures and heart rate variability (HRV) in individuals with (CAC+) and without CAC (CAC-). Methods Hemodynamic measures from 56 healthy, middle-aged (51.0 [43.0-58.0] years [median, (Q1-Q3)]) individuals (41 men/15 women) participating in a 91-km (251.2 [217.2-271.6] minutes) leisure sport mountain bike race were included in this study. 25 participants (20 men/5 women), were classified as CAC+ based on coronary computer tomographic assessment. Hemodynamic measures and HRV were used to assess the study subjects at the top of the hardest hill (THH) during the last quarter of the race. Results At the top of THH, CAC+ individuals had significantly higher systolic blood pressure (235.0 [225.0-245.0] mmHg vs 220.0 [192.5-237.5] mmHg, p=0.008), higher diastolic blood pressure (DBP) (105.0 [95.0-110.0] mmHg vs 95.0 [85.0-110.0] mmHg, p=0.006), higher pulse pressure (130.0 [125.0-140.0] mmHg vs 123.0 [110.0-130.0] mmHg, p=0.039), higher mean rate pressure product (33882 [30872-35053] bpm x mmHg vs 31028 [27392-33047] bpm x mmHg, p=0.028), and larger increase in DBP from baseline (20.0 [20.0-30.0] mmHg vs 10.0 [0.0-20.0] mmHg, p=0.001), compared with CAC- individuals. Further, CAC+ showed a significant reduction in the low-frequency component of HRV (HRVLF) (6.3 [2.4-11.5] ms2 vs 12.4 [6.8-20.2] ms2, p=0.044). In multivariate analysis, HRVLF was an independent predictor of presence of CAC even after adjusting for established risk factors of atherosclerosis. Conclusion CAC+ individuals had significant alterations in hemodynamic measures and HRVLF following prolonged high-intensity exercise compared to individuals without CAC. HRVLF was an independent predictor of CAC, suggesting an adverse autonomic response to high-intensity exercise in individuals with CAC.

Artificial intelligence systems show promise to aid in the di- agnostic pathway of prostate cancer (PC), by supporting radiologists in interpreting magnetic resonance images (MRI) of the prostate. Most MRI-based systems are designed to detect clinically significant PC le- sions, with the main objective of preventing over-diagnosis. Typically, these systems involve an automatic prostate segmentation component and a clinically significant PC lesion detection component. In spite of the compound nature of the systems, evaluations are presented assum- ing a standalone clinically significant PC detection component. That is, they are evaluated in an idealized scenario and under the assumption that a highly accurate prostate segmentation is available at test time. In this work, we aim to evaluate a clinically significant PC lesion de- tection system accounting for its compound nature. For that purpose, we simulate a realistic deployment scenario and evaluate the effect of two non-ideal and previously validated prostate segmentation modules on the PC detection ability of the compound system. Following, we com- pare them with an idealized setting, where prostate segmentations are assumed to have no faults. We observe significant differences in the de- tection ability of the compound system in a realistic scenario and in the presence of the highest-performing prostate segmentation module (DSC: 90.07+-0.74), when compared to the idealized one (AUC: 77.93 +- 3.06 and 84.30+- 4.07, P<.001). Our results depict the relevance of holistic evalu- ations for PC detection compound systems, where interactions between system components can lead to decreased performance and degradation at deployment time.

Early prediction of Alzheimer’s Disease (AD) is widely addressed within dementia research, where the use of deep learning (DL) has received increased popularity across disciplines within healthcare, particularly medical imaging. While unintentional bias in data collection and study design in the field of DL has been a highly debated topic, systematic analysis of the performance of DL models for female and male sub-groups is rarely considered during development. Recently, some studies have shed light on the importance of analyzing classification results for uneven performance with respect to protected characteristics like gender and ethnicity. However, the results are inconsistent and heterogeneous, especially for gender fairness analysis in DL classification of AD. In our work, we seek to present a standardized analysis of gender fairness in the AD classification scenario and discuss the lack of a benchmark to evaluate skewed performance with respect to gender for DL and AD studies. For that purpose, we train two deep neural networks to classify between cognitively normal (CN) and AD subjects from magnetic resonance imaging (MRI) and analyze the classification ability for male and female sub-groups separately. Both networks achieve high classification performance at testing time with an area under the receiver operating characteristics (ROC) curve (AUC) of 1) 0.97±0.05 and 2) 0.96±0.03. Contrary to other studies, our sub-group analyses do not show significant differences in performance between male and female subgroups.

The most prevalent form of bladder cancer is urothelial carcinoma, characterized by a high recurrence rate and substantial lifetime treatment costs for patients. Grading is a prime factor for patient risk stratification, although it suffers from inconsistencies and variations among pathologists. Moreover, absence of annotations in medical imaging renders it difficult to train deep learning models. To address these challenges, we introduce a pipeline designed for bladder cancer grading using histological slides. First, it extracts urothelium tissue tiles at different magnification levels, employing a convolutional neural network for processing for feature extraction. Then, it engages in the slide-level prediction process. It employs a nested multiple-instance learning approach with attention to predict the grade. To distinguish different levels of malignancy within specific regions of the slide, we include the origins of the tiles in our analysis. The attention scores at region level are shown to correlate with verified high-grade regions, giving some explainability to the model. Clinical evaluations demonstrate that our model consistently outperforms previous state-of-the-art methods, achieving an F1 score of 0.85.

Introduction: Patients who regain return of spontaneous circulation (ROSC) after in-hospital cardiac arrest are often critically ill and at risk of re-arrest. However, re-arrest is insufficiently studied. Pre-hospital data indicate a re-arrest rate ranging from 3% to 39%. Our study aims to assess the immediate hazard of re-arrest after ROSC, depending on whether the patient’s last observed rhythm before ROSC was shockable or not. Methods: We analyzed defibrillator recordings and clinical data from 763 cardiac arrest episodes at four different hospitals. ROSC was defined as an organized ECG rhythm compatible with a pulse, accompanied by the absence of chest compressions for at least one minute. An organized rhythm with a QRS frequency ≥ 12 was categorized as pulseless electrical activity (PEA). Conversely, a QRS frequency < 12 or a flat line represented asystole. Ventricular fibrillation or tachycardia (VF/VT) was identified based on its distinct morphology. We further stratified ROSC based on whether the preceding rhythm was shockable or not. After comparing four different parametric time-to-event models, we chose the most useful one and estimated the immediate hazard of re-arrest along the timeline of resuscitation. Results: After the initial event of cardiac arrest, we observed 316 re-arrests. Among these, 68% relapsed to PEA, 25% relapsed to VF/VT, and 7% relapsed to asystole. Summarized in the figure, the initial hazard of re-arrest from ROSC after PEA or asystole to a non-shockable rhythm was 0.02 per minute. By the 9th minute, this hazard decreased to 0.01 per minute. Meanwhile, the hazard for re-arrest to a shockable rhythm remained constant at 0.01 per minute. For re-arrest from ROSC after VF/VT back to VF/VT, the hazard was 0.05 per minute initially, decreasing to 0.03 per minute by the 12th minute. The corresponding hazard for re-arrest to PEA or asystole remained at 0.01 per minute. Conclusion: The hazard of re-arrest after return of spontaneous circulation (ROSC) to either pulseless electrical activity (PEA), asystole, or ventricular fibrillation/tachycardia (VF/VT) varies by the last observed state before ROSC. Notably, re-arrest to VF/VT following ROSC after previous VF/VT poses the highest risk. This understanding can assist healthcare professionals in anticipating events during the critical minutes following successful resuscitation and adjusting treatment accordingly.

Patients undergoing cardiopulmonary ressuscitation (CPR) may respond through rhythm transitions between different rhythms ventricular fibrillation (VF), ventricular tachycardia (VT), asystole (AS), pulseless electrical activity (PEA) and pulse generating rhythm (PR). Rhythm recognition is crucial to address adequate resuscitation efforts, and in this study we applied a deep neural network to classify ECG rhythms during cardiac arrest. Artifact-free four second segments were extracted from 100 patients in out-of-hospital cardiac arrest. A convolutional neural network (CNN) was trained to discriminate between five cardiac arrest rhytm types. Experiments were conducted with increasing number of layers. For each model, training was repeated 10 times to explore variations in the results. A five layer network provided the best performance with an accuracy of 80. 3 (78.1,81.3)% (median(25th, 75th quartiles)). We have proposed a deep learning approach to automatically recognise five cardiac arrest rhythms common during resuscitation.

Traditional deep learning (DL) approaches based on supervised learning paradigms require large amounts of annotated data that are rarely available in the medical domain. Unsupervised Out-of-distribution (OOD) detection is an alternative that requires less annotated data. Further, OOD applications exploit the class skewness commonly present in medical data. Magnetic resonance imaging (MRI) has proven to be useful for prostate cancer (PCa) diagnosis and management, but current DL approaches rely on T2w axial MRI, which suffers from low out-of-plane resolution. We propose a multi-stream approach to accommodate different T2w directions to improve the performance of PCa lesion detection in an OOD approach. We evaluate our approach on a publicly available data-set, obtaining better detection results in terms of AUC when compared to a single direction approach (73.1 vs 82.3). Our results show the potential of OOD approaches for PCa lesion detection based on MRI.

Background Scar size is critical to left ventricular (LV) remodeling and adverse outcomes following myocardial infarction (MI). Late Gadolinium-enhancement (LGE) in cardiac magnetic resonance imaging is the gold standard for assessing MI size. Texture-based probability mapping (TPM) is a novel machine learning-based analysis of LGE images. This proof-of-concept study investigates the potential clinical implications of temporal changes in TPM during the first year following an acute revascularized MI. Methods 41 patients with first-time acute ST-elevation MI were included in this study. All patients had a single-vessel disease and were successfully revascularized by primary percutaneous coronary intervention. LGE images were obtained two days, one week, two months, and one year post-MI. MI size by TPM was compared with manual LGE-based MI calculation, LV remodeling, and biomarkers. Results TPM showed a significant increase in infarct size from the second month through the first year (p<0.01). MI size estimated by TPM at all different time points demonstrated strong correlations with peak Troponin T levels. At one week, TPM assessment correlated positively with maximum C-reactive protein (r=0.54, p<0.01), and at two months, TPM positively correlated with N-Terminal Pro Brain Natriuretic Peptide. Conclusion This proof-of-concept study suggests that TPM may provide additional information to conventional LGE-based MI analysis of scar formation, LV remodeling, and biomarkers following an acute revascularized MI. Highlights ### Competing Interest Statement The authors have declared no competing interest. ### Funding Statement This work was supported by Helse Vest grant no. 912296 ### Author Declarations I confirm all relevant ethical guidelines have been followed, and any necessary IRB and/or ethics committee approvals have been obtained. Yes The details of the IRB/oversight body that provided approval or exemption for the research described are given below: The study was approved by the Regional Ethics Committee at the University of Bergen and was conducted according to the Declaration of Helsinki principles. All patients gave informed consent before their inclusion in the study. I confirm that all necessary patient/participant consent has been obtained and the appropriate institutional forms have been archived, and that any patient/participant/sample identifiers included were not known to anyone (e.g., hospital staff, patients or participants themselves) outside the research group so cannot be used to identify individuals. Yes I understand that all clinical trials and any other prospective interventional studies must be registered with an ICMJE-approved registry, such as ClinicalTrials.gov. I confirm that any such study reported in the manuscript has been registered and the trial registration ID is provided (note: if posting a prospective study registered retrospectively, please provide a statement in the trial ID field explaining why the study was not registered in advance). Yes I have followed all appropriate research reporting guidelines, such as any relevant EQUATOR Network research reporting checklist(s) and other pertinent material, if applicable. Yes The data that support the findings of this study are available on request from the corresponding author, [V.F] * CVD : cardiovascular disease MI : Myocardial infarction STEMI : ST-elevation myocardial infarction MVO : microvascular obstruction TPM : Texture-based probability mapping SI : Signal intensity CMR : Cardiac magnetic resonance imaging LGE-CMR : late gadolinium-enhanced cardiac magnetic resonance imaging PCI : percutaneous coronary intervention. LV : Left ventricular LVEF : Left ventricular ejection fraction LVEDVi : Left ventricular end-diastolic volume index LVESVi : Left ventricular end-systolic volume index eGFR : estimated glomerular filtration rate hs-CRP : high-sensitivity C-reactive protein NT-proBNP : N-terminal pro-B-type natriuretic peptide TnT : Cardiac troponin-T

Newborn mortality is a global challenge with around 2.4 million neonatal deaths in 2019. One third of these occur within the first-and-only day of life with labour complications and birth asphyxia being the primary causes. Existing guidelines for newborn resuscitation are based on limited scientific evidence, and evidens based research is sought for. To increase our knowledge on resuscitation of newborns, it is crucial to first quantify what is currently being done in terms of therapeutic activities, such as ventilation and stimulation, and how they affect resuscitation outcomes. In the current study, the therapeutic activities during newborn resuscitation are quantified by estimating a timeline describing the start and stop of activities. The proposed approach is combining methods using both video and time series data recorded during resuscitation, where the predictions are based on the available sources. From video the activity recognition is done by a 3D CNN method. For the signal data feature extraction is performed on ECG and accelerometer signals and thereafter machine learning is done to perform stimulation detection. We show that best results are achieved with all signals and video available, for the activity "stimulation"we get an AUC of 0.86, sensitivity of 82.32%, specificity of 82.23%, and precision of 57.59%. If only signals or video is available we still get good results with AUC at 0.80, and 0.84 respectively.