Integrating novel medical concepts and relationships into existing ontologies can significantly enhance their coverage and utility for both biomedical research and clinical applications. Clinical notes, as unstructured documents rich with detailed patient observations, offer valuable context-specific insights and represent a promising yet underutilized source for ontology extension. Despite this potential, directly leveraging clinical notes for ontology extension remains largely unexplored. To address this gap, we propose CLOZE, a novel framework that uses large language models (LLMs) to automatically extract medical entities from clinical notes and integrate them into hierarchical medical ontologies. By capitalizing on the strong language understanding and extensive biomedical knowledge of pre-trained LLMs, CLOZE effectively identifies disease-related concepts and captures complex hierarchical relationships. The zero-shot framework requires no additional training or labeled data, making it a cost-efficient solution. Furthermore, CLOZE ensures patient privacy through automated removal of protected health information (PHI). Experimental results demonstrate that CLOZE provides an accurate, scalable, and privacy-preserving ontology extension framework, with strong potential to support a wide range of downstream applications in biomedical research and clinical informatics.

When applying offline reinforcement learning (RL) in healthcare scenarios, the out-of-distribution (OOD) issues pose significant risks, as inappropriate generalization beyond clinical expertise can result in potentially harmful recommendations. While existing methods like conservative Q-learning (CQL) attempt to address the OOD issue, their effectiveness is limited by only constraining action selection by suppressing uncertain actions. This action-only regularization imitates clinician actions that prioritize short-term rewards, but it fails to regulate downstream state trajectories, thereby limiting the discovery of improved long-term treatment strategies. To safely improve policy beyond clinician recommendations while ensuring that state-action trajectories remain in-distribution, we propose \textit{Offline Guarded Safe Reinforcement Learning} ($\mathsf{OGSRL}$), a theoretically grounded model-based offline RL framework. $\mathsf{OGSRL}$ introduces a novel dual constraint mechanism for improving policy with reliability and safety. First, the OOD guardian is established to specify clinically validated regions for safe policy exploration. By constraining optimization within these regions, it enables the reliable exploration of treatment strategies that outperform clinician behavior by leveraging the full patient state history, without drifting into unsupported state-action trajectories. Second, we introduce a safety cost constraint that encodes medical knowledge about physiological safety boundaries, providing domain-specific safeguards even in areas where training data might contain potentially unsafe interventions. Notably, we provide theoretical guarantees on safety and near-optimality: policies that satisfy these constraints remain in safe and reliable regions and achieve performance close to the best possible policy supported by the data. When evaluated on the MIMIC-III sepsis treatment dataset, $\mathsf{OGSRL}$ demonstrated significantly better OOD handling than baselines. $\mathsf{OGSRL}$ achieved a 78\% reduction in mortality estimates and a 51\% increase in reward compared to clinician decisions.

Linking (aligning) biomedical concepts across diverse data sources enables various integrative analyses, but it is challenging due to the discrepancies in concept naming conventions. Various strategies have been developed to overcome this challenge, such as those based on string-matching rules, manually crafted thesauri, and machine learning models. However, these methods are constrained by limited prior biomedical knowledge and can hardly generalize beyond the limited amounts of rules, thesauri, or training samples. Recently, large language models (LLMs) have exhibited impressive results in diverse biomedical NLP tasks due to their unprecedentedly rich prior knowledge and strong zero-shot prediction abilities. However, LLMs suffer from issues including high costs, limited context length, and unreliable predictions. In this research, we propose PromptLink, a novel biomedical concept linking framework that leverages LLMs. It first employs a biomedical-specialized pre-trained language model to generate candidate concepts that can fit in the LLM context windows. Then it utilizes an LLM to link concepts through two-stage prompts, where the first-stage prompt aims to elicit the biomedical prior knowledge from the LLM for the concept linking task and the second-stage prompt enforces the LLM to reflect on its own predictions to further enhance their reliability. Empirical results on the concept linking task between two EHR datasets and an external biomedical KG demonstrate the effectiveness of PromptLink. Furthermore, PromptLink is a generic framework without reliance on additional prior knowledge, context, or training data, making it well-suited for concept linking across various types of data sources. The source code of this study is available at https://github.com/constantjxyz/PromptLink.

Atrial fibrillation (AF) is a common cardiac arrhythmia with serious health consequences if not detected and treated early. Detecting AF using wearable devices with photoplethysmography (PPG) sensors and deep neural networks has demonstrated some success using proprietary algorithms in commercial solutions. However, to improve continuous AF detection in ambulatory settings towards a population-wide screening use case, we face several challenges, one of which is the lack of large-scale labeled training data. To address this challenge, we propose to leverage AF alarms from bedside patient monitors to label concurrent PPG signals, resulting in the largest PPG-AF dataset so far (8.5 M 30-second records from 24,100 patients) and demonstrating a practical approach to build large labeled PPG datasets. Furthermore, we recognize that the AF labels thus obtained contain errors because of false AF alarms generated from imperfect built-in algorithms from bedside monitors. Dealing with label noise with unknown distribution characteristics in this case requires advanced algorithms. We, therefore, introduce and open-source a novel loss design, the cluster membership consistency (CMC) loss, to mitigate label errors. By comparing CMC with state-of-the-art methods selected from a noisy label competition, we demonstrate its superiority in handling label noise in PPG data, resilience to poor-quality signals, and computational efficiency.

PPG signal is a valuable resource for continuous heart rate monitoring; however, this signal suffers from artifact movements, which is particularly relevant during physical exercise and makes this biomedical signal difficult to use for heart rate detection during those activities. The purpose of this study was to develop learning models to determine heart rate using data from wearables (PPG and acceleration signals) and dealing with noise during physical exercise. Learning models based on CNNs and LSTMs were developed to predict the heart rate. The PPG signal was combined with data from accelerometers trying to overcome the noise movement on the PPG signal. Two datasets were used on this work: the 2015 IEEE Signal Processing Cup (SPC) dataset was used for training and testing, and another dataset was used for validation of the learning model (PPG-DaLiA dataset). The predictions obtained by the learning model represented a mean average error of 7.033±5.376 bpm for the SCP dataset, while a mean average error of 9.520±8.443 bpm for the validation set. The use of acceleration data increases the performance of the learning models on the prediction of the heart rate, showing the benefits of using this source of data to overcome the noise movement problem on the PPG signal. The combination of PPG signal with acceleration data could allow the learning models to use more information regarding the motion artifacts that affect the PPG and improve performance on the physiological event detections, which will largely spread the use of wearables on the healthcare applications for continuous monitor the physiological state allowing early and accurate detection of pathological events.

Personalized diagnosis prediction based on electronic health records (EHR) of patients is a promising yet challenging task for AI in healthcare. Existing studies typically ignore the heterogeneity of diseases across different patients. For example, diabetes can have different complications across different patients (e.g., hyperlipidemia and circulatory disorder), which requires personalized diagnoses and treatments. Specifically, existing models fail to consider 1) varying severity of the same diseases for different patients, 2) complex interactions among syndromic diseases, and 3) dynamic progression of chronic diseases. In this work, we propose to perform personalized diagnosis prediction based on EHR data via capturing disease severity, interaction, and progression. In particular, we enable personalized disease representations via severity-driven embeddings at the disease level. Then, at the visit level, we propose to capture higher-order interactions among diseases that can collectively affect patients’ health status via hypergraph-based aggregation; at the patient level, we devise a personalized generative model based on neural ordinary differential equations to capture the continuous-time disease progressions underlying discrete and incomplete visits. Extensive experiments on two real-world EHR datasets show significant performance gains brought by our approach, yielding average improvements of 10.70% for diagnosis prediction over state-of-the-art competitors.

Photoplethysmography (PPG) is a ubiquitous physiological measurement that detects beat-to-beat pulsatile blood volume changes and hence has a potential for monitoring cardiovascular conditions, particularly in ambulatory settings. A PPG dataset that is created for a particular use case is often imbalanced, due to a low prevalence of the pathological condition it targets to predict and the paroxysmal nature of the condition as well. To tackle this problem, we propose log-spectral matching GAN (LSM-GAN), a generative model that can be used as a data augmentation technique to alleviate the class imbalance in a PPG dataset to train a classifier. LSM-GAN utilizes a novel generator that generates a synthetic signal without a up-sampling process of input white noises, as well as adds the mismatch between real and synthetic signals in frequency domain to the conventional adversarial loss. In this study, experiments are designed focusing on examining how the influence of LSM-GAN as a data augmentation technique on one specific classification task - atrial fibrillation (AF) detection using PPG. We show that by taking spectral information into consideration, LSM-GAN as a data augmentation solution can generate more realistic PPG signals.

Early detection of Atrial Fibrillation (AFib) is crucial to prevent stroke recurrence. New tools for monitoring cardiac rhythm are important for risk stratification and stroke prevention. As many of new approaches to long-term AFib detection are now based on photoplethysmogram (PPG) recordings from wearable devices, ensuring high PPG signal-to-noise ratios is a fundamental requirement for a robust detection of AFib episodes. Traditionally, signal quality assessment is often based on the evaluation of similarity between pulses to derive signal quality indices. There are limitations to using this approach for accurate assessment of PPG quality in the presence of arrhythmia, as in the case of AFib, mainly due to substantial changes in pulse morphology. In this paper, we first tested the performance of algorithms selected from a body of studies on PPG quality assessment using a dataset of PPG recordings from patients with AFib. We then propose machine learning approaches for PPG quality assessment in 30-s segments of PPG recording from 13 stroke patients admitted to the University of California San Francisco (UCSF) neuro intensive care unit and another dataset of 3764 patients from one of the five UCSF general intensive care units. We used data acquired from two systems, fingertip PPG (fPPG) from a bedside monitor system, and radial PPG (rPPG) measured using a wearable commercial wristband. We compared various supervised machine learning techniques including k-nearest neighbors, decisions trees, and a two-class support vector machine (SVM). SVM provided the best performance. fPPG signals were used to build the model and achieved 0.9477 accuracy when tested on the data from the fPPG exclusive to the test set, and 0.9589 accuracy when tested on the rPPG data.

We propose a projection method based on singular value decomposition (SVD) to validate arterial blood pressure (ABP) signal in order to avoid artifacts and noise in subsequent processing. The projection has been done on 567 validated ABP beats collected from 51 patients hospitalized in University of California, Los Angeles Medical Center. Then, we compare the performance of the proposed projection method with that of a previously developed algorithm, signal abnormality index (SAI), which is a value- and trend-based approach, and has shown to be effective in cleaning the ABP waveforms. The testing dataset consists of 1336 ten-second ABP segments (18 472 ABP beats) of both valid and invalid pulses selected randomly from multiparameter intelligent monitoring for intensive care II database. The proposed projection approach that validates the signal based on the shape of the waveform achieves a true positive rate (TPR) of 99.06%, 5.43% higher than that of the SAI, and a false positive rate (FPR) of 7.69%, 17.38% lower than that of SAI. Integration of some of the SAI-value-based abnormality conditions to the validation process of SVD-based method can further improve the performance by reducing the FPR to 3.92%, while keeping the TPR at the high rate of 99.05%.

The only established technique for intracranial pressure (ICP) measurement is an invasive procedure requiring surgically penetrating the skull for placing pressure sensors. However, there are many clinical scenarios where a noninvasive assessment of ICP is highly desirable. With an assumption of a linear relationship among arterial blood pressure (ABP), ICP, and flow velocity (FV) of major cerebral arteries, an approach has been previously developed to estimate ICP noninvasively, the core of which is the linear estimation of the coefficients f between ABP and ICP from the coefficients w calculated between ABP and FV. In this paper, motivated by the fact that the relationships among these three signals are so complex that simple linear models may be not adequate to depict the relationship between these two coefficients, i.e., f and w , we investigate the adoption of several nonlinear kernel regression approaches, including kernel spectral regression (KSR) and support vector machine (SVM) to improve the original linear ICP estimation approach. The ICP estimation results on a dataset consisting of 446 entries from 23 patients show that the mean ICP error by the nonlinear approaches can be reduced to below 6.0 mmHg compared to 6.7 mmHg of the original approach. The statistical test also demonstrates that the ICP error by the proposed nonlinear kernel approaches is statistically smaller than that estimated with the original linear model (p < 0.05). The current result confirms the potential of using nonlinear regression to achieve more accurate noninvasive ICP assessment.