The field of healthcare has increasingly turned its focus towards Large Language Models (LLMs) due to their remarkable performance. However, their performance in actual clinical applications has been underexplored. Traditional evaluations based on question-answering tasks don't fully capture the nuanced contexts. This gap highlights the need for more in-depth and practical assessments of LLMs in real-world healthcare settings. Objective: We sought to evaluate the performance of LLMs in the complex clinical context of adult critical care medicine using systematic and comprehensible analytic methods, including clinician annotation and adjudication. Methods: We investigated the performance of three general LLMs in understanding and processing real-world clinical notes. Concepts from 150 clinical notes were identified by MetaMap and then labeled by 9 clinicians. Each LLM's proficiency was evaluated by identifying the temporality and negation of these concepts using different prompts for an in-depth analysis. Results: GPT-4 showed overall superior performance compared to other LLMs. In contrast, both GPT-3.5 and text-davinci-003 exhibit enhanced performance when the appropriate prompting strategies are employed. The GPT family models have demonstrated considerable efficiency, evidenced by their cost-effectiveness and time-saving capabilities. Conclusion: A comprehensive qualitative performance evaluation framework for LLMs is developed and operationalized. This framework goes beyond singular performance aspects. With expert annotations, this methodology not only validates LLMs' capabilities in processing complex medical data but also establishes a benchmark for future LLM evaluations across specialized domains.
This paper explores the challenges in evaluating machine learning (ML) models for continuous health monitoring using wearable devices beyond conventional metrics. We state the complexities posed by real-world variability, disease dynamics, user-specific characteristics, and the prevalence of false notifications, necessitating novel evaluation strategies. Drawing insights from large-scale heart studies, the paper offers a comprehensive guideline for robust ML model evaluation on continuous health monitoring.
Atrial fibrillation (AF) is a prevalent cardiac arrhythmia associated with significant health ramifications, including an elevated susceptibility to ischemic stroke, heart disease, and heightened mortality. Photoplethysmography (PPG) has emerged as a promising technology for continuous AF monitoring for its cost-effectiveness and widespread integration into wearable devices. Our team previously conducted an exhaustive review on PPG-based AF detection before June 2019. However, since then, more advanced technologies have emerged in this field. This paper offers a comprehensive review of the latest advancements in PPG-based AF detection, utilizing digital health and artificial intelligence (AI) solutions, within the timeframe spanning from July 2019 to December 2022. Through extensive exploration of scientific databases, we have identified 59 pertinent studies. Our comprehensive review encompasses an in-depth assessment of the statistical methodologies, traditional machine learning techniques, and deep learning approaches employed in these studies. In addition, we address the challenges encountered in the domain of PPG-based AF detection. Furthermore, we maintain a dedicated website to curate the latest research in this area, with regular updates on a regular basis.
Atrial fibrillation (AF) is the most common type of cardiac arrhythmia. It is associated with an increased risk of stroke, heart failure, and other cardiovascular complications, but can be clinically silent. Passive AF monitoring with wearables may help reduce adverse clinical outcomes related to AF. Detecting AF in noisy wearable data poses a significant challenge, leading to the emergence of various deep learning techniques. Previous deep learning models learn from a single modality, either electrocardiogram (ECG) or photoplethysmography (PPG) signals. However, deep learning models often struggle to learn generalizable features and rely on features that are more susceptible to corruption from noise, leading to sub-optimal performances in certain scenarios, especially with low-quality signals. Given the increasing availability of ECG and PPG signal pairs from wearables and bedside monitors, we propose a new approach, SiamAF, leveraging a novel Siamese network architecture and joint learning loss function to learn shared information from both ECG and PPG signals. At inference time, the proposed model is able to predict AF from either PPG or ECG and outperforms baseline methods on three external test sets. It learns medically relevant features as a result of our novel architecture design. The proposed model also achieves comparable performance to traditional learning regimes while requiring much fewer training labels, providing a potential approach to reduce future reliance on manual labeling.
Smart watches and other wearable devices are equipped with photoplethysmography (PPG) sensors for monitoring heart rate and other aspects of cardiovascular health. However, PPG signals collected from such devices are susceptible to corruption from noise and motion artifacts, which cause errors in heart rate estimation. Typical denoising approaches filter or reconstruct the signal in ways that eliminate much of the morphological information, even from the clean parts of the signal that would be useful to preserve. In this work, we develop an algorithm for denoising PPG signals that reconstructs the corrupted parts of the signal, while preserving the clean parts of the PPG signal. Our novel framework relies on self-supervised training, where we leverage a large database of clean PPG signals to train a denoising autoencoder. As we show, our reconstructed signals provide better estimates of heart rate from PPG signals than the leading heart rate estimation methods. Further experiments show significant improvement in Heart Rate Variability (HRV) estimation from PPG signals using our algorithm. We conclude that our algorithm denoises PPG signals in a way that can improve downstream analysis of many different health metrics from wearable devices.
Photoplethysmography (PPG) provides a low-cost, non-invasive method to continuously monitor various cardiovascular parameters. PPG signals are generated by wearable devices and frequently contain large artifacts caused by external factors, such as motion of the human subject. In order to ensure robust and accurate extraction of physiological parameters, corrupted areas of the signal need to be identified and handled appropriately. Previous methodology relied either on handcrafted feature detectors or signal metrics which yield sub-optimal performance, or relied on machine learning techniques such as deep neural networks (DNN) which lack interpretability and are computationally and memory intensive. In this work, we present a novel method to learn a small set of interpretable convolutional kernels that has performance similar to -- and often better than -- the state-of-the-art DNN approach with several orders of magnitude fewer parameters. This work allows for efficient, robust, and interpretable signal quality assessment and artifact segmentation on low-power devices.
Obtaining large-scale well-annotated is always a daunting challenge, especially in the medical research domain because of the shortage of domain expert. Instead of human annotation, in this work, we use the alarm information generated from bed-side monitor to get the pseudo label for the co-current photoplethysmography (PPG) signal. Based on this strategy, we end up with over 8 million 30-second PPG segment. To solve the label noise caused by false alarms, we propose the cluster consistency, which use an unsupervised auto-encoder (hence not subject to label noise) approach to cluster training samples into a finite number of clusters. Then the learned cluster membership is used in the subsequent supervised learning phase to force the distance in the latent space of samples in the same cluster to be small while that of samples in different clusters to be big. In the experiment, we compare with the state-of-the-art algorithms and test on external datasets. The results show the superiority of our method in both classification performance and efficiency.
Training machine learning algorithms from a small and imbalanced dataset is often a daunting challenge in medical research. However, it has been shown that the synthetic data generated by data augmentation techniques can enlarge the dataset and contribute to alleviating the imbalance situation. In this study, we propose a novel generative adversarial network (GAN) architecture-Welch-GAN and focused on examining how its influence on classifier performance is related to signal quality and class imbalance within the context of photoplethysmography (PPG)-based atrial fibrillation (AF) detection. Pulse oximetry data were collected from 126 adult patients and augmented using the permutation technique to build a large training set for training an AF detection model based on a one-dimensional residual neural network. To test the model, PPG data were collected from 13 stroke patients and utilized. Four data augmentation methods, including both traditional and GANs, are leveraged as baseline in this study. Three different experiments are designed to investigate each data augmentation methods from the aspect of performance gain, robustness to motion artifact and training sample size, respectively. Compared to the un-augmented data, by training the same AF classification algorithm using augmented data, the AF detection accuracy was significantly improved from 80.36% to over 90% with no compromise on sensitivity nor on negative predicted value. Within each data augmentation techniques, Welch-GAN has shown around 3% superiority in terms of AF detection accuracy compared to the baseline methods, which suggests the state-of-the-art of our proposed Welch-GAN.
As one of the prevalent topic mining tools, neural topic modeling has attracted a lot of interests for the advantages of high efficiency in training and strong generalisation abilities. However, due to the lack of context in each short text, the existing neural topic models may suffer from feature sparsity on such documents. To alleviate this issue, we propose a Context Reinforced Neural Topic Model (CRNTM), whose characteristics can be summarized as follows. Firstly, by assuming that each short text covers only a few salient topics, CRNTM infers the topic for each word in a narrow range. Secondly, our model exploits pre-trained word embeddings by treating topics as multivariate Gaussian distributions or Gaussian mixture distributions in the embedding space. Extensive experiments on two benchmark datasets validate the effectiveness of the proposed model on both topic discovery and text classification.