Improving Multimodal Learning with Multi-Loss Gradient Modulation

Learning from multiple modalities, such as audio and video, offers opportunities for leveraging complementary information, enhancing robustness, and improving contextual understanding and performance. However, combining such modalities presents challenges, especially when modalities differ in data structure, predictive contribution, and the complexity of their learning processes. It has been observed that one modality can potentially dominate the learning process, hindering the effective utilization of information from other modalities and leading to sub-optimal model performance. To address this issue the vast majority of previous works suggest to assess the unimodal contributions and dynamically adjust the training to equalize them. We improve upon previous work by introducing a multi-loss objective and further refining the balancing process, allowing it to dynamically adjust the learning pace of each modality in both directions, acceleration and deceleration, with the ability to phase out balancing effects upon convergence. We achieve superior results across three audio-video datasets: on CREMA-D, models with ResNet backbone encoders surpass the previous best by 1.9% to 12.4%, and Conformer backbone models deliver improvements ranging from 2.8% to 14.1% across different fusion methods. On AVE, improvements range from 2.7% to 7.7%, while on UCF101, gains reach up to 6.1%.

Multimodal wearable EEG, EMG and accelerometry measurements improve the accuracy of tonic-clonic seizure detection in-hospital

Objective: Most current wearable tonic-clonic seizure (TCS) detection systems are based on extra-cerebral signals, such as electromyography (EMG) or accelerometry (ACC). Although many of these devices show good sensitivity in seizure detection, their false positive rates (FPR) are still relatively high. Wearable EEG may improve performance; however, studies investigating this remain scarce. This paper aims 1) to investigate the possibility of detecting TCSs with a behind-the-ear, two-channel wearable EEG, and 2) to evaluate the added value of wearable EEG to other non-EEG modalities in multimodal TCS detection. Method: We included 27 participants with a total of 44 TCSs from the European multicenter study SeizeIT2. The multimodal wearable detection system Sensor Dot (Byteflies) was used to measure two-channel, behind-the-ear EEG, EMG, electrocardiography (ECG), ACC and gyroscope (GYR). First, we evaluated automatic unimodal detection of TCSs, using performance metrics such as sensitivity, precision, FPR and F1-score. Secondly, we fused the different modalities and again assessed performance. Algorithm-labeled segments were then provided to a neurologist and a wearable data expert, who reviewed and annotated the true positive TCSs, and discarded false positives (FPs). Results: Wearable EEG outperformed the other modalities in unimodal TCS detection by achieving a sensitivity of 100.0% and a FPR of 10.3/24h (compared to 97.7% sensitivity and 30.9/24h FPR for EMG; 95.5% sensitivity and 13.9 FPR for ACC). The combination of wearable EEG and EMG achieved overall the most clinically useful performance in offline TCS detection with a sensitivity of 97.7%, a FPR of 0.4/24 h, a precision of 43.0%, and a F1-score of 59.7%. Subsequent visual review of the automated detections resulted in maximal sensitivity and zero FPs.

Increasing Performance And Sample Efficiency With Model-agnostic Interactive Feature Attributions

Model-agnostic feature attributions can provide local insights in complex ML models. If the explanation is correct, a domain expert can validate and trust the model's decision. However, if it contradicts the expert's knowledge, related work only corrects irrelevant features to improve the model. To allow for unlimited interaction, in this paper we provide model-agnostic implementations for two popular explanation methods (Occlusion and Shapley values) to enforce entirely different attributions in the complex model. For a particular set of samples, we use the corrected feature attributions to generate extra local data, which is used to retrain the model to have the right explanation for the samples. Through simulated and real data experiments on a variety of models we show how our proposed approach can significantly improve the model's performance only by augmenting its training dataset based on corrected explanations. Adding our interactive explanations to active learning settings increases the sample efficiency significantly and outperforms existing explanatory interactive strategies. Additionally we explore how a domain expert can provide feature attributions which are sufficiently correct to improve the model.

Explaining the Model and Feature Dependencies by Decomposition of the Shapley Value

Shapley values have become one of the go-to methods to explain complex models to end-users. They provide a model agnostic post-hoc explanation with foundations in game theory: what is the worth of a player (in machine learning, a feature value) in the objective function (the output of the complex machine learning model). One downside is that they always require outputs of the model when some features are missing. These are usually computed by taking the expectation over the missing features. This however introduces a non-trivial choice: do we condition on the unknown features or not? In this paper we examine this question and claim that they represent two different explanations which are valid for different end-users: one that explains the model and one that explains the model combined with the feature dependencies in the data. We propose a new algorithmic approach to combine both explanations, removing the burden of choice and enhancing the explanatory power of Shapley values, and show that it achieves intuitive results on simple problems. We apply our method to two real-world datasets and discuss the explanations. Finally, we demonstrate how our method is either equivalent or superior to state-to-of-art Shapley value implementations while simultaneously allowing for increased insight into the model-data structure.

U-PASS: an Uncertainty-guided deep learning Pipeline for Automated Sleep Staging

As machine learning becomes increasingly prevalent in critical fields such as healthcare, ensuring the safety and reliability of machine learning systems becomes paramount. A key component of reliability is the ability to estimate uncertainty, which enables the identification of areas of high and low confidence and helps to minimize the risk of error. In this study, we propose a machine learning pipeline called U-PASS tailored for clinical applications that incorporates uncertainty estimation at every stage of the process, including data acquisition, training, and model deployment. The training process is divided into a supervised pre-training step and a semi-supervised finetuning step. We apply our uncertainty-guided deep learning pipeline to the challenging problem of sleep staging and demonstrate that it systematically improves performance at every stage. By optimizing the training dataset, actively seeking informative samples, and deferring the most uncertain samples to an expert, we achieve an expert-level accuracy of 85% on a challenging clinical dataset of elderly sleep apnea patients, representing a significant improvement over the baseline accuracy of 75%. U-PASS represents a promising approach to incorporating uncertainty estimation into machine learning pipelines, thereby improving their reliability and unlocking their potential in clinical settings.

CoRe-Sleep: A Multimodal Fusion Framework for Time Series Robust to Imperfect Modalities

Sleep abnormalities can have severe health consequences. Automated sleep staging, i.e. labelling the sequence of sleep stages from the patient's physiological recordings, could simplify the diagnostic process. Previous work on automated sleep staging has achieved great results, mainly relying on the EEG signal. However, often multiple sources of information are available beyond EEG. This can be particularly beneficial when the EEG recordings are noisy or even missing completely. In this paper, we propose CoRe-Sleep, a Coordinated Representation multimodal fusion network that is particularly focused on improving the robustness of signal analysis on imperfect data. We demonstrate how appropriately handling multimodal information can be the key to achieving such robustness. CoRe-Sleep tolerates noisy or missing modalities segments, allowing training on incomplete data. Additionally, it shows state-of-the-art performance when testing on both multimodal and unimodal data using a single model on SHHS-1, the largest publicly available study that includes sleep stage labels. The results indicate that training the model on multimodal data does positively influence performance when tested on unimodal data. This work aims at bridging the gap between automated analysis tools and their clinical utility.

Automated Movement Detection with Dirichlet Process Mixture Models and Electromyography

Numerous sleep disorders are characterised by movement during sleep, these include rapid-eye movement sleep behaviour disorder (RBD) and periodic limb movement disorder. The process of diagnosing movement related sleep disorders requires laborious and time-consuming visual analysis of sleep recordings. This process involves sleep clinicians visually inspecting electromyogram (EMG) signals to identify abnormal movements. The distribution of characteristics that represent movement can be diverse and varied, ranging from brief moments of tensing to violent outbursts. This study proposes a framework for automated limb-movement detection by fusing data from two EMG sensors (from the left and right limb) through a Dirichlet process mixture model. Several features are extracted from 10 second mini-epochs, where each mini-epoch has been classified as 'leg-movement' or 'no leg-movement' based on annotations of movement from sleep clinicians. The distributions of the features from each category can be estimated accurately using Gaussian mixture models with the Dirichlet process as a prior. The available dataset includes 36 participants that have all been diagnosed with RBD. The performance of this framework was evaluated by a 10-fold cross validation scheme (participant independent). The study was compared to a random forest model and outperformed it with a mean accuracy, sensitivity, and specificity of 94\%, 48\%, and 95\%, respectively. These results demonstrate the ability of this framework to automate the detection of limb movement for the potential application of assisting clinical diagnosis and decision-making.

L-SeqSleepNet: Whole-cycle Long Sequence Modelling for Automatic Sleep Staging

Human sleep is cyclical with a period of approximately 90 minutes, implying long temporal dependency in the sleep data. Yet, exploring this long-term dependency when developing sleep staging models has remained untouched. In this work, we show that while encoding the logic of a whole sleep cycle is crucial to improve sleep staging performance, the sequential modelling approach in existing state-of-the-art deep learning models are inefficient for that purpose. We thus introduce a method for efficient long sequence modelling and propose a new deep learning model, L-SeqSleepNet, which takes into account whole-cycle sleep information for sleep staging. Evaluating L-SeqSleepNet on four distinct databases of various sizes, we demonstrate state-of-the-art performance obtained by the model over three different EEG setups, including scalp EEG in conventional Polysomnography (PSG), in-ear EEG, and around-the-ear EEG (cEEGrid), even with a single EEG channel input. Our analyses also show that L-SeqSleepNet is able to alleviate the predominance of N2 sleep (the major class in terms of classification) to bring down errors in other sleep stages. Moreover the network becomes much more robust, meaning that for all subjects where the baseline method had exceptionally poor performance, their performance are improved significantly. Finally, the computation time only grows at a sub-linear rate when the sequence length increases.

EventNet: Detecting Events in EEG

Neurologists are often looking for various "events of interest" when analyzing EEG. To support them in this task various machine-learning-based algorithms have been developed. Most of these algorithms treat the problem as classification, thereby independently processing signal segments and ignoring temporal dependencies inherent to events of varying duration. At inference time, the predicted labels for each segment then have to be post processed to detect the actual events. We propose an end-to-end event detection approach (EventNet), based on deep learning, that directly works with events as learning targets, stepping away from ad-hoc postprocessing schemes to turn model outputs into events. We compare EventNet with a state-of-the-art approach for artefact and and epileptic seizure detection, two event types with highly variable durations. EventNet shows improved performance in detecting both event types. These results show the power of treating events as direct learning targets, instead of using ad-hoc postprocessing to obtain them. Our event detection framework can easily be extended to other event detection problems in signal processing, since the deep learning backbone does not depend on any task-specific features.

Feature matching as improved transfer learning technique for wearable EEG

Objective: With the rapid rise of wearable sleep monitoring devices with non-conventional electrode configurations, there is a need for automated algorithms that can perform sleep staging on configurations with small amounts of labeled data. Transfer learning has the ability to adapt neural network weights from a source modality (e.g. standard electrode configuration) to a new target modality (e.g. non-conventional electrode configuration). Methods: We propose feature matching, a new transfer learning strategy as an alternative to the commonly used finetuning approach. This method consists of training a model with larger amounts of data from the source modality and few paired samples of source and target modality. For those paired samples, the model extracts features of the target modality, matching these to the features from the corresponding samples of the source modality. Results: We compare feature matching to finetuning for three different target domains, with two different neural network architectures, and with varying amounts of training data. Particularly on small cohorts (i.e. 2 - 5 labeled recordings in the non-conventional recording setting), feature matching systematically outperforms finetuning with mean relative differences in accuracy ranging from 0.4% to 4.7% for the different scenarios and datasets. Conclusion: Our findings suggest that feature matching outperforms finetuning as a transfer learning approach, especially in very low data regimes. Significance: As such, we conclude that feature matching is a promising new method for wearable sleep staging with novel devices.