Developing new machine learning applications often requires the collection of new datasets. However, existing datasets may already contain relevant information to train models for new purposes. We propose SoundCollage: a framework to discover new classes within audio datasets by incorporating (1) an audio pre-processing pipeline to decompose different sounds in audio samples, and (2) an automated model-based annotation mechanism to identify the discovered classes. Furthermore, we introduce the clarity measure to assess the coherence of the discovered classes for better training new downstream applications. Our evaluations show that the accuracy of downstream audio classifiers within discovered class samples and a held-out dataset improves over the baseline by up to 34.7% and 4.5%, respectively. These results highlight the potential of SoundCollage in making datasets reusable by labeling with newly discovered classes. To encourage further research in this area, we open-source our code at github.com/nokia-bell-labs/audio-class-discovery.

Photoplethysmography (PPG) is the leading non-invasive technique for monitoring biosignals and cardiovascular health, with widespread adoption in both clinical settings and consumer wearable devices. While machine learning models trained on PPG signals have shown promise, they tend to be task-specific and struggle with generalization. Current research is limited by the use of single-device datasets, insufficient exploration of out-of-domain generalization, and a lack of publicly available models, which hampers reproducibility. To address these limitations, we present PaPaGei, the first open foundation model for PPG signals. The model is pre-trained on over 57,000 hours of data, comprising 20 million unlabeled PPG segments from publicly available datasets. We introduce a novel representation learning approach that leverages domain knowledge of PPG signal morphology across individuals, enabling the capture of richer representations compared to traditional contrastive learning methods. We evaluate PaPaGei against state-of-the-art time-series foundation models and self-supervised learning benchmarks across 20 tasks from 10 diverse datasets, spanning cardiovascular health, sleep disorders, pregnancy monitoring, and wellbeing assessment. Our model demonstrates superior performance, improving classification and regression metrics by 6.3% and 2.9% respectively in at least 14 tasks. Notably, PaPaGei achieves these results while being more data- and parameter-efficient, outperforming models that are 70x larger. Beyond accuracy, we examine model robustness across different skin tones, establishing a benchmark for bias evaluation in future models. PaPaGei can serve as both a feature extractor and an encoder for multimodal models, opening up new opportunities for multimodal health monitoring. Models, data, and code are available at: https://github.com/nokia-bell-labs/papagei-foundation-model

Contrastive learning (CL) has emerged as a promising approach for representation learning in time series data by embedding similar pairs closely while distancing dissimilar ones. However, existing CL methods often introduce false negative pairs (FNPs) by neglecting inherent characteristics and then randomly selecting distinct segments as dissimilar pairs, leading to erroneous representation learning, reduced model performance, and overall inefficiency. To address these issues, we systematically define and categorize FNPs in time series into semantic false negative pairs and temporal false negative pairs for the first time: the former arising from overlooking similarities in label categories, which correlates with similarities in non-stationarity and the latter from neglecting temporal proximity. Moreover, we introduce StatioCL, a novel CL framework that captures non-stationarity and temporal dependency to mitigate both FNPs and rectify the inaccuracies in learned representations. By interpreting and differentiating non-stationary states, which reflect the correlation between trends or temporal dynamics with underlying data patterns, StatioCL effectively captures the semantic characteristics and eliminates semantic FNPs. Simultaneously, StatioCL establishes fine-grained similarity levels based on temporal dependencies to capture varying temporal proximity between segments and to mitigate temporal FNPs. Evaluated on real-world benchmark time series classification datasets, StatioCL demonstrates a substantial improvement over state-of-the-art CL methods, achieving a 2.9% increase in Recall and a 19.2% reduction in FNPs. Most importantly, StatioCL also shows enhanced data efficiency and robustness against label scarcity.

Self-supervised learning (SSL) has become the de facto training paradigm of large models, where pre-training is followed by supervised fine-tuning using domain-specific data and labels. Despite demonstrating comparable performance with supervised methods, comprehensive efforts to assess SSL's impact on machine learning fairness (i.e., performing equally on different demographic breakdowns) are lacking. Hypothesizing that SSL models would learn more generic, hence less biased representations, this study explores the impact of pre-training and fine-tuning strategies on fairness. We introduce a fairness assessment framework for SSL, comprising five stages: defining dataset requirements, pre-training, fine-tuning with gradual unfreezing, assessing representation similarity conditioned on demographics, and establishing domain-specific evaluation processes. We evaluate our method's generalizability on three real-world human-centric datasets (i.e., MIMIC, MESA, and GLOBEM) by systematically comparing hundreds of SSL and fine-tuned models on various dimensions spanning from the intermediate representations to appropriate evaluation metrics. Our findings demonstrate that SSL can significantly improve model fairness, while maintaining performance on par with supervised methods-exhibiting up to a 30% increase in fairness with minimal loss in performance through self-supervision. We posit that such differences can be attributed to representation dissimilarities found between the best- and the worst-performing demographics across models-up to x13 greater for protected attributes with larger performance discrepancies between segments. Code: https://github.com/Nokia-Bell-Labs/SSLfairness

Self-supervised learning (SSL) has shown remarkable performance in computer vision tasks when trained offline. However, in a Continual Learning (CL) scenario where new data is introduced progressively, models still suffer from catastrophic forgetting. Retraining a model from scratch to adapt to newly generated data is time-consuming and inefficient. Previous approaches suggested re-purposing self-supervised objectives with knowledge distillation to mitigate forgetting across tasks, assuming that labels from all tasks are available during fine-tuning. In this paper, we generalize self-supervised continual learning in a practical setting where available labels can be leveraged in any step of the SSL process. With an increasing number of continual tasks, this offers more flexibility in the pre-training and fine-tuning phases. With Kaizen1, we introduce a training architecture that is able to mitigate catastrophic forgetting for both the feature extractor and classifier with a carefully designed loss function. By using a set of comprehensive evaluation metrics reflecting different aspects of continual learning, we demonstrated that Kaizen significantly outperforms previous SSL models in competitive vision benchmarks, with up to 16.5% accuracy improvement on split CIFAR-100. Kaizen is able to balance the trade-off between knowledge retention and learning from new data with an end-to-end model, paving the way for practical deployment of continual learning systems.

Limited availability of labeled data for machine learning on multimodal time-series extensively hampers progress in the field. Self-supervised learning (SSL) is a promising approach to learn data representations without relying on labels. However, existing SSL methods require expensive computations of negative pairs and are typically designed for single modalities, which limits their versatility. We introduce CroSSL (Cross-modal SSL), which puts forward two novel concepts: masking intermediate embeddings produced by modality-specific encoders, and their aggregation into a global embedding through a cross-modal aggregator CroSSL allows for handling missing modalities and end-to-end cross-modal earning without requiring prior data preprocessing for handling missing inputs or negative-pair sampling for contrastive learning. We evaluate our method on a wide range of data, including motion sensors such as accelerometers or gyroscopes and biosignals (heart rate, electroencephalograms, electromyograms, electrooculograms, and electrodermal). Overall, CroSSL outperforms previous SSL and supervised benchmarks using minimal labeled data, and also sheds light on how latent masking can improve cross-modal learning.

OBJECTIVES: Large language models (LLMs) have demonstrated remarkable generalization and across diverse tasks, leading individuals to increasingly use them as personal assistants due to their emerging reasoning capabilities. Nevertheless, a notable obstacle emerges when including numerical/temporal data into these prompts, such as data sourced from wearables or electronic health records. LLMs employ tokenizers in their input that break down text into smaller units. However, tokenizers are not designed to represent numerical values and might struggle to understand repetitive patterns and context, treating consecutive values as separate tokens and disregarding their temporal relationships. This article discusses the challenges of representing and tokenizing temporal data. It argues that naively passing timeseries to LLMs can be ineffective due to the modality gap between numbers and text. MATERIALS AND METHODS: We conduct a case study by tokenizing a sample mobile sensing dataset using the OpenAI tokenizer. We also review recent works that feed timeseries data into LLMs for human-centric tasks, outlining common experimental setups like zero-shot prompting and few-shot learning. RESULTS: The case study shows that popular LLMs split timestamps and sensor values into multiple nonmeaningful tokens, indicating they struggle with temporal data. We find that preliminary works rely heavily on prompt engineering and timeseries aggregation to "ground" LLMs, hinting that the "modality gap" hampers progress. The literature was critically analyzed through the lens of models optimizing for expressiveness versus parameter efficiency. On one end of the spectrum, training large domain-specific models from scratch is expressive but not parameter-efficient. On the other end, zero-shot prompting of LLMs is parameter-efficient but lacks expressiveness for temporal data. DISCUSSION: We argue tokenizers are not optimized for numerical data, while the scarcity of timeseries examples in training corpora exacerbates difficulties. We advocate balancing model expressiveness and computational efficiency when integrating temporal data. Prompt tuning, model grafting, and improved tokenizers are highlighted as promising directions. CONCLUSION: We underscore that despite promising capabilities, LLMs cannot meaningfully process temporal data unless the input representation is addressed. We argue that this paradigm shift in how we leverage pretrained models will particularly affect the area of biomedical signals, given the lack of modality-specific foundation models.

Time series forecasting, as one of the fundamental machine learning areas, has attracted tremendous attentions over recent years. The solutions have evolved from statistical machine learning (ML) methods to deep learning techniques. One emerging sub-field of time series forecasting is individual disease progression forecasting, e.g., predicting individuals' disease development over a few days (e.g., deteriorating trends, recovery speed) based on few past observations. Despite the promises in the existing ML techniques, a variety of unique challenges emerge for disease progression forecasting, such as irregularly-sampled time series, data sparsity, and individual heterogeneity in disease progression. To tackle these challenges, we propose novel Conditional Neural Ordinary Differential Equations Processes (CNDPs), and validate it in a COVID-19 disease progression forecasting task using audio data. CNDPs allow for irregularly-sampled time series modelling, enable accurate forecasting with sparse past observations, and achieve individual-level progression forecasting. CNDPs show strong performance with an Unweighted Average Recall (UAR) of 78.1%, outperforming a variety of commonly used Recurrent Neural Networks based models. With the proposed label-enhancing mechanism (i.e., including the initial health status as input) and the customised individual-level loss, CNDPs further boost the performance reaching a UAR of 93.6%. Additional analysis also reveals the model's capability in tracking individual-specific recovery trend, implying the potential usage of the model for remote disease progression monitoring. In general, CNDPs pave new pathways for time series forecasting, and provide considerable advantages for disease progression monitoring.

The development of fast and accurate screening tools, which could facilitate testing and prevent more costly clinical tests, is key to the current pandemic of COVID-19. In this context, some initial work shows promise in detecting diagnostic signals of COVID-19 from audio sounds. In this paper, we propose a voice-based framework to automatically detect individuals who have tested positive for COVID-19. We evaluate the performance of the proposed framework on a subset of data crowdsourced from our app, containing 828 samples from 343 participants. By combining voice signals and reported symptoms, an AUC of 0.79 has been attained, with a sensitivity of 0.68 and a specificity of 0.82. We hope that this study opens the door to rapid, low-cost, and convenient pre-screening tools to automatically detect the disease.

The INTERSPEECH 2021 Computational Paralinguistics Challenge addresses four different problems for the first time in a research competition under well-defined conditions: In the COVID-19 Cough and COVID-19 Speech Sub-Challenges, a binary classification on COVID-19 infection has to be made based on coughing sounds and speech; in the Escalation SubChallenge, a three-way assessment of the level of escalation in a dialogue is featured; and in the Primates Sub-Challenge, four species vs background need to be classified. We describe the Sub-Challenges, baseline feature extraction, and classifiers based on the 'usual' COMPARE and BoAW features as well as deep unsupervised representation learning using the AuDeep toolkit, and deep feature extraction from pre-trained CNNs using the Deep Spectrum toolkit; in addition, we add deep end-to-end sequential modelling, and partially linguistic analysis.

Body-focused repetitive behaviors (BFRBs), like face-touching or skin-picking, are hand-driven behaviors which can damage one’s appearance, if not identified early and treated. Technology for automatic detection is still under-explored, with few previous works being limited to wearables with single modalities (e.g., motion). Here, we propose a multi-sensory approach combining motion, orientation, and heart rate sensors to detect BFRBs. We conducted a feasibility study in which participants (N=10) were exposed to BFRBs-inducing tasks, and analyzed 380 mins of signals1 under an extensive evaluation of sensing modalities, cross-validation methods, and observation windows. Our models achieved an AUC > 0.90 in distinguishing BFRBs, which were more evident in observation windows 5 mins prior to the behavior as opposed to 1-min ones. In a follow-up qualitative survey, we found that not only the timing of detection matters but also models need to be context-aware, when designing just-in-time interventions to prevent BFRBs.

Despite advances in hardware and software enabling faster on-device inference, training Deep Neural Networks (DNN) models has largely been a long-running task over TBs of collected user data in centralised repositories. Federated Learning has emerged as an alternative, privacy-preserving paradigm to train models without accessing directly on-device data, by leveraging device resources to create per client updates and aggregate centrally. This has been applied to various tasks, ranging from next-word prediction to automatic speech recognition (ASR). In this tutorial, we recognise on-device sensing as a privacy-sensitive task and build a federated learning system from scratch to showcase how to train a model for accelerometer-based activity recognition in a federated manner. In addition, we present the current landscape and challenges in the realm of federated learning and mobile sensing and provide guidelines on how to build such systems in a privacy-preserving and scalable manner.

In this perspective we want to highlight the rise of what we call "digital phenotyping" or inferring insights about peopleãs health and behavior from their digital devices and data, and the challenges this introduces. Indeed, the collection, processing, and storage of data comes with significant ethical, security and data governance considerations. The COVID-19 pandemic has laid bare the importance of scientific data and modeling, both to understand the nature and spread of the disease, and to develop treatment. But digital devices have also played a (controversial) role, with track and trace systems and increasingly "vaccine passports" being rolled out to help societies open back up. These systems epitomize a wider and longer-standing trend towards seeing almost any form of personal data as potentially health data, especially with the rise of consumer health trackers and other gadgets. Here, we offer an overview of the risks this introduces, drawing on the earlier revolution in genomic sequencing, and propose guidelines to help protect privacy whilst utilizing personal data to help get society back up to speed.

Journal of the American Medical Informatics Association, Volume 28, Issue 9, https://doi.org/10.1093/jamia/ocab012 In the originally published version of this manuscript, the wrong article accepted date, 21 January 2020, was recorded. The paper has been corrected to reflect the correct article accepted date, 21 January 2021.

Audio signals generated by the human body (e.g., sighs, breathing, heart, digestion, vibration sounds) have routinely been used by clinicians as indicators to diagnose disease or assess disease progression. Until recently, such signals were usually collected through manual auscultation at scheduled visits. Research has now started to use digital technology to gather bodily sounds (e.g., from digital stethoscopes) for cardiovascular or respiratory examination, which could then be used for automatic analysis. Some initial work shows promise in detecting diagnostic signals of COVID-19 from voice and coughs. In this paper we describe our data analysis over a large-scale crowdsourced dataset of respiratory sounds collected to aid diagnosis of COVID-19. We use coughs and breathing to understand how discernible COVID-19 sounds are from those in asthma or healthy controls. Our results show that even a simple binary machine learning classifier is able to classify correctly healthy and COVID-19 sounds. We also show how we distinguish a user who tested positive for COVID-19 and has a cough from a healthy user with a cough, and users who tested positive for COVID-19 and have a cough from users with asthma and a cough. Our models achieve an AUC of above 80% across all tasks. These results are preliminary and only scratch the surface of the potential of this type of data and audio-based machine learning. This work opens the door to further investigation of how automatically analysed respiratory patterns could be used as pre-screening signals to aid COVID-19 diagnosis.

Smartphones have started to be used as self reporting tools for mental health state as they accompany individuals during their days and can therefore gather temporally fine grained data. However, the analysis of self reported mood data offers challenges related to non-homogeneity of mood assessment among individuals due to the complexity of the feeling and the reporting scales, as well as the noise and sparseness of the reports when collected in the wild. In this paper, we propose a new end-to-end ML model inspired by video frame prediction and machine translation, that forecasts future sequences of mood from previous self-reported moods collected in the real world using mobile devices. Contrary to traditional time series forecasting algorithms, our multi-task encoder-decoder recurrent neural network learns patterns from different users, allowing and improving the prediction for users with limited number of self-reports. Unlike traditional feature-based machine learning algorithms, the encoder-decoder architecture enables to forecast a sequence of future moods rather than one single step. Meanwhile, multi-task learning exploits some unique characteristics of the data (mood is bi-dimensional), achieving better results than when training single-task networks or other classifiers.