publications

Cardiac arrest (CA) may cause severe brain damage, cognitive impairments and death. Monitoring neurological recovery after hospitalization is critical to provide suitable treatment. In this study, we aim to develop an algorithm to aid in neurological recovery classification. The dataset used for this purpose includes 1020 patients and contains both continuous sensor measurements, taken from 0 to 72 hours after the CA, and structured data. More specifically the dataset is split into a training set (60%), validation set (10%) and undisclosed test set (30%). The developed model uses a one-dimensional convolutional neural network to extract features from 5-minute time series segments, fed into a recurrent neural network, to capture temporal information and provide adaptability in recording length. The output features are then merged with embedded patient metadata in a fully connected layer, for the final classification of neurological outcome. The project is part of the George B. Moody PhysioNet Challenge 2023, where our team (EEG-Attackers) achieved a challenge score and rank on the undisclosed test dataset of 0.16 (30), 0.13 (33), 0.11 (34) and 0.12 (35) using recordings from the first 12, 24, 48 and 72 hours after CA.

Medical data such as the electrocardiogram (ECG) has received an increased interest within biometric settings. One of the main benefits is the difficulty in counterfeiting the information due to its hidden nature. However, medical information may be exposed to intra-subject variability. This is evident in extraction of soft biometric traits from ECG, where results can vary widely depending on cardiac condition and status. This work investigates methods of lowering the variability, by employing multi-task learning on a shared feature extractor. Three different architectures are suggested and benchmarked. Specifically, experiments are carried out on age and gender estimation showing state of the art results on two public ECG datasets, while lowering estimation variance among instances.

Measurement of the left ventricular outflow tract diameter (LVOTd) in echocardiography is a common source of error when used to calculate the stroke volume. The aim of this study is to assess whether a deep learning (DL) model, trained on a clinical echocardiographic dataset, can perform automatic LVOTd measurements on par with expert cardiologists.

This challenge aims to reduce CO2 emissions from Norwegian construction machines by utilizing data-driven approaches. Skanska Norge AS provides a wealth of data from their construction sites, including GPS data, machine data, vibration data, and drone maps. The challenge focuses on using this data to optimize road construction processes for increased sustainability. Participants are tasked with developing feasible solutions to reduce idle time, optimize dump truck flow, minimize unnecessary driving, and improve driving styles. The evaluation criteria consider the novelty of the algorithms, feasibility of implementation, and the sustainable impact. The competition encourages innovative approaches to leverage data for sustainable construction practices.

The Assisted Cement Log Interpretation Project has used machine learning (ML) to create a tool that interprets cement logs by predicting a predefined set of annular condition codes used in the cement log interpretation process.The development of a cement log interpretation tool speeds up the log interpretation process and enables expert knowledge to be efficiently shared when training new professionals. By using high-quality and consistent training data sets, the project has trained a model that will support unbiased and consistent interpretations over time.The tool consists of a training and a prediction tool integrated with cased-hole logging interpretation software. By containerizing the code using an “API First” design principle (API: application programming interface), the applicability of this add-on tool is broad. The ML model is trained using selected and engineered features from cement logs, and the tool predicts an annular condition code according to the cement classification system for each depth segment in the log. The interpreters can easily fetch a complete cement log interpretation prediction for the log and use that as a template for their final interpretation. The ML model can easily be retrained with new data sets to improve accuracy even further.To improve cement log interpretation consistency in the industry, the code will be made available as open source.

Electrocardiography is one of the most frequently used methods to evaluate cardiovascular diseases. However, the last decade has shown that deep convolutional neural networks (CNN) can extract information from the electrocardiogram (ECG) that goes beyond traditional diagnostics, such as predicting a persons age. In this study, we trained two different 1-dimensional CNNs on open datasets to predict age from a persons ECG. The models were trained and validated using 10 seconds long 12-lead ECG records, resampled to 100Hz. 59355 ECGs were used for training and cross-validation, while 21748 ECGs from a separate cohort were used as the test set. We compared the performance achieved on the cross-validation with the performance on the test set. Furthermore, we used cardiologist annotated cardiovascular conditions to categorize the patients in the test set in order to assess whether some cardiac condition leads to greater discrepancies between CNN-predicted age and chronological age. The best CNN model, using an Inception Time architecture, showed a significant drop in performance, in terms of mean absolute error (MAE), from cross-validation on the training set (7.90 ± 0.04 years) to the performance on the test set (8.3 years). On the other hand, the mean squared error (MSE) improved from the training set (117.5 ± 2.7 years^2) to the test set (111 years^2). We also observed that the cardiovascular condition that showed the highest deviation between predicted and biological age, in terms of MAE, was the patients with pacing rhythm (10.5 years), while the patients with prolonged QT-interval had the smallest deviation (7.4 years) in terms of MAE. This work contributes to existing knowledge of age prediction using deep CNNs on ECGs by showing how a trained model performs on a test set from a separate cohort to that used in the training set.

Murmurs are sounds caused by turbulent blood flow that are often the first sign of structural heart disease. These sounds are detected by auscultating the heart using a stethoscope, or more recently by a phonocardiogram (PCG). We aim to identify the presence, absence, or un-clear cases of murmurs, as well as predict normal or abnormal clinical outcome from PCG recordings using machine learning. We trained and tested two 1-dimensional convolutional neural networks (CNN) on a PCG data set from a pediatric population of 1568 individuals. One model predicted mur-murs, while the other model predicted clinical outcomes. Both models were trained to give recording-wise predictions, while the final predictions were given for every patient (patient wise predictions). This paper describes our participation in the George B. Moody PhysioNet Challenge 2022 whose objective was to identify heart murmurs and clinical outcome from PCGs. Our team, Simulab, trained a clinical outcome classifier that achieved a challenge cost score of 12419 (ranked 14th out of 39 teams) and the murmur classifier achieved a weighted accuracy of 0.593 (ranked 30th out of 40 teams) on the test set.

Athletes often have training-induced remodeling of the heart, and this can sometimes be seen as abnormal but non-pathological changes in the electrocardiogram. However, these changes can be confused with severe cardiovascular diseases that, in some cases, can cause cardiovascular death. Electrocardiogram data from athletes is therefore important to learn more about the difference between normal athletic remodeling and pathological remodeling of the heart. This work provides a dataset of electrocardiograms from 28 Norwegian elite endurance athletes. The electrocardiograms are standard 12-lead resting ECGs, recorded for 10 seconds while the athlete’s lay supine on a bench. The electrocardiograms were then interpreted by an interpretation algorithm and by a trained cardiologist. The electrocardiogram waveform data and the interpretations were stored in Python Waveform Database format and made publicly available through PhysioNet.

Immature cardiomyocytes, such as those obtained by stem cell differentiation, have been shown to be useful alternatives to mature cardiomyocytes, which are limited in availability and difficult to obtain, for evaluating the behaviour of drugs for treating arrhythmia. In silico models of induced pluripotent stem cell-derived cardiomyocytes (iPSC-CMs) can be used to simulate the behaviour of the transmembrane potential and cytosolic calcium under drug-treated conditions. Simulating the change in action potentials due to various ionic current blocks enables the approximation of drug behaviour. We used eight machine learning classification models to predict partial block of seven possible ion currents \ (}textit{I}_{}textit{CaL}},}textit{I}_{}textit{Kr}},}textit{I}_{}textit{to}},}textit{I}_{}textit{K1}},}textit{I}_{}textit{Na}},}textit{I}_{}textit{NaL}} and }textit{I}_{}textit{Ks}}) \ in a simulated dataset containing nearly 4600 action potentials represented as a paired measure of transmembrane potential and cytosolic calcium. Each action potential was generated under 1 \ }textit{H}_{}textit{z}} \ pacing. The Convolutional Neural Network outperformed the other models with an average accuracy of predicting partial ionic current block of 93% in noise-free data and 72% accuracy with 3% added random noise. Our results show that \ }textit{I}_{}textit{CaL}} \ and \ }textit{I}_{}textit{Kr}} \ current block were classified with high accuracy with and without noise. The classification of \ }textit{I}_{}textit{to}} \ , \ }textit{I}_{}textit{K1}} \ and \ }textit{I}_{}textit{Na}} \ current block showed high accuracy at 0% noise, but showed a significant decrease in accuracy when noise was added. Finally, the accuracy of \ }textit{I}_{}textit{NaL}} \ and \ }textit{I}_{}textit{Ks}} \ classification were relatively lower than the other current blocks at 0% noise and also showed a significant drop in accuracy when noise was added. In conclusion, these machine learning methods may present a pathway for estimating drug response in adult phenotype cardiac systems, but the data must be sufficiently filtered to remove noise before being used with classifier algorithms.

The electrocardiogram (ECG) is an almost universally accessible diagnostic tool for heart disease. An ECG is measured by using an electrocardiograph, and today’s electrocardiographs use built-in software to interpret the ECGs automatically after they are recorded. However, these algorithms show limited performance, and therefore clinicians usually have to manually interpret the ECG, regardless of whether an algorithm has interpreted the ECG or not. Manual interpretation of the ECG can be time-consuming and require specific skills. Therefore, a better algorithm is clearly needed to make correct ECG interpretations more accessible and time efficient. Algorithms based on artificial intelligence have shown promising performance in many fields, including ECG interpretation, over the last few years and might represent an alternative to manual ECG interpretation.

Due to the possibilities in miniaturization and wearability, photoplethysmography (PPG) has recently gained a large interest not only for heart rate measurement, but also for estimating heart rate variability, which is derived from ECG by convention. The agreement between PPG and ECG-based HRV has been assessed in several studies, but the feasibility of PPG-based HRV estimation is still largely unknown for many conditions. In this study, we assess the feasibility of HRV estimation based on finger PPG during rest, mild physical exercise and mild mental stress. In addition, we compare different variants of signal processing methods including selection of fiducial point and outlier correction. Based on five minutes synchronous recordings of PPG and ECG from 15 healthy participants during each of these three conditions, the PPG-based HRV estimation was assessed for the SDNN and RMSSD parameters, calculated based on two different fiducial points (foot point and maximum slope), with and without outlier correction. The results show that HRV estimation based on finger PPG is feasible during rest and mild mental stress, but can give large errors during mild physical exercise. A good estimation is very dependent on outlier correction and fiducial point selection, and SDNN seems to be a more robust parameter compared to RMSSD for PPG-based HRV estimation.

Over the last decade, AI has shown its feasibility in classifying heart-related diagnoses from ECGs. Earlier studies have mainly focused on 12 and 2-lead ECGs, but we aim to classify 26 different diagnoses based on 12, 6, 4, 3, and 2-lead ECGs in this study. We trained a supervised model on a dataset containing 88 253 ECGs with 26 different diagnoses used as ground truth. The training and classification steps can be separated into three parts. (1) Pan Tompkins algorithm was used to find peaks and calculate the average heart rate. (2) The average heart rate and the Fourier transformed ECG signal was used to train convolutional neural networks (CNN) system that classified the ECGs with regular or irregular rhythms. 9 out of 26 classes were classified in this step. (3) Finally, CNN models in a classifier chain were trained to classify the remaining 17 diagnoses. The classification results from step 2 and the raw ECG signal were used as input to the classifier chain in step 3. Our team, CardiOUS, achieved a PhysioNet Challenge score of −0.63 for all sets of leads on the hidden test set. Based on the test score, our team placed at 38th out of 39 teams in the official ranking.

Colorectal cancer is one of the deadliest and most widespread types of cancer in the world. Colonoscopy is the procedure used to detect and diagnose polyps from the colon, but today’s detection rate shows a significant error rate that affects diagnosis and treatment. An automatic image segmentation algorithm may help doctors to improve the detection rate of pathological polyps in the colon. Furthermore, segmenting endoscopic tools in images taken during colonoscopy may contribute towards robotic assisted surgery. In this study, we trained and validated both pre-trained and not pre-trained segmentation models on two different data sets, containing images of polyps and endoscopic tools. Finally, we applied the models on two separate test sets and the best polyp model got a dice score 0.857 and the test instrument model got a dice score 0.948. Moreover, we found that pre-training of the models increased the performance in segmenting polyps and endoscopic tools.

MedAI: Transparency in Medical Image Segmentation is a challenge held for the first time at the Nordic AI Meet that focuses on medical image segmentation and transparency in machine learning (ML)-based systems. We propose three tasks to meet specific gastrointestinal image segmentation challenges collected from experts within the field, including two separate segmentation scenarios and one scenario on transparent ML systems. The latter emphasizes the need for explainable and interpretable ML algorithms. We provide a development dataset for the participants to train their ML models, tested on a concealed test dataset.

The objective of this study was to classify 27 cardiac abnormalities based on a data set of 43101 ECG recordings. A hybrid model combining a rule-based algorithm with different deep learning architectures was developed. We compared two different Convolutional Neural Networks, a Fully Convolutional Neural Network and an Encoder Network, a combination of both, and with the addition of another neural network using age and gender as input. Two of these combinations were finally combined with a rule-based model using derived ECG features. The performance of the models was evaluated on validation data during model development using hold-out validation. Finally, the models were deployed to a Docker image, trained on the provided development data, and tested on the Challenge validation set. The model that performed best on the Challenge validation set was then deployed and tested on the full Challenge test set. The performance was evaluated based on a particular Challenge score. Our team, TeamUIO, achieved a Challenge validation score of 0.377, and a full test score of 0.206 for our best model. The score on the full test set placed us at 20th out of 41 teams in the official ranking.