Department of Computer Science

Generative Artificial Intelligence is set to revolutionize healthcare delivery by transforming traditional patient care into a more personalized, efficient, and proactive process. Chatbots, serving as interactive conversational models, will probably drive this patient-centered transformation in healthcare. Through the provision of various services, including diagnosis, personalized lifestyle recommendations, dynamic scheduling of follow-ups, and mental health support, the objective is to substantially augment patient health outcomes, all the while mitigating the workload burden on healthcare providers. The life-critical nature of healthcare applications necessitates establishing a unified and comprehensive set of evaluation metrics for conversational models. Existing evaluation metrics proposed for various generic large language models (LLMs) demonstrate a lack of comprehension regarding medical and health concepts and their significance in promoting patients’ well-being. Moreover, these metrics neglect pivotal user-centered aspects, including trust-building, ethics, personalization, empathy, user comprehension, and emotional support. The purpose of this paper is to explore state-of-the-art LLM-based evaluation metrics that are specifically applicable to the assessment of interactive conversational models in healthcare. Subsequently, we present a comprehensive set of evaluation metrics designed to thoroughly assess the performance of healthcare chatbots from an end-user perspective. These metrics encompass an evaluation of language processing abilities, impact on real-world clinical tasks, and effectiveness in user-interactive conversations. Finally, we engage in a discussion concerning the challenges associated with defining and implementing these metrics, with particular emphasis on confounding factors such as the target audience, evaluation methods, and prompt techniques involved in the evaluation process.

This dataset was collected from university students before, during, and after the COVID-19 lockdown in Southern California. Data collection happened continuously for the average of 7.8 months (SD=3.8, MIN=1.0, MAX=13.4) from a population of 21 students of which 12 have also completed an exit survey, and 7 started before the California COVID-19 lockdown order. This multimodal dataset included different means of data collection such as Samsung Galaxy Watch, Oura Ring, a Life-logger app named Personicle, a questionnaire mobile app named Personicle Questions, and periodical and personalised surveys. The dataset contains raw data from Photoplethysmogram (PPG), Inertial measurement unit (IMU), and pressure sensors in addition to processed data on heart rate, heart rate variability, sleep (bedtime, sleep stages, quality), and physical activity (step, active calories, type of activity). Ecological momentary assessments were collected from participants on daily and weekly bases containing their Positive and Negative Affect Schedule (PANAS) questionnaire and their emotional responses to COVID-19 and their health. Subjective data was also collected through monthly surveys containing standard mood and mental health surveys such as Beck Depression Inventory II (BDI-II), Brief Symptom Inventory (BSI), GAD-7, Inclusion of Other in the Self Scale (IOS-Partner), Acceptability of Intervention Measure (AIM), Intervention Appropriateness Measure (IAM), Feasibility of Intervention Measure (FIM), Experiences in Close Relationships Scale Short Form (ECR-S), UCLA Three-Item Loneliness Scale (ULS), Multidimensional Scale of Perceived Social Support (MSPSS), Investment Model Scale (IMS), Conflict Management Scale (CMS), etc in addition to their response to important events and COVID-19. This dataset can be used to study emotions, mood, physical activity, and lifestyle of young adults through longitudinal subjective and objective measures. This dataset also contains valuable data regarding adjustment of lifestyle and emotions during the events of 2020 and 2021 including COVID-19 discovery and lockdown, Black Life Matter movement, 2020 US presidential elections, etc. On average, participants engaged in the EMA collection study at a rate of 86% (SD=10, MIN=65, MAX=99). Smartwatch usage saw an average participation rate of 51% (SD=20, MIN=16, MAX=88), while engagement with the Oura ring averaged at 85% (SD=12, MIN=60, MAX=99).

The profound impact of food on health necessitates advanced nutrition-oriented food recommendation services. Conventional methods often lack the crucial elements of personalization, explainability, and interactivity. While Large Language Models (LLMs) bring interpretability and explainability, their standalone use falls short of achieving true personalization. In this paper, we introduce ChatDiet, a novel LLM-powered framework designed specifically for personalized nutrition-oriented food recommendation chatbots. ChatDiet integrates personal and population models, complemented by an orchestrator, to seamlessly retrieve and process pertinent information. The personal model leverages causal discovery and inference techniques to assess personalized nutritional effects for a specific user, whereas the population model provides generalized information on food nutritional content. The orchestrator retrieves, synergizes and delivers the output of both models to the LLM, providing tailored food recommendations designed to support targeted health outcomes. The result is a dynamic delivery of personalized and explainable food recommendations, tailored to individual user preferences. Our evaluation of ChatDiet includes a compelling case study, where we establish a causal personal model to estimate individual nutrition effects. Our assessments, including a food recommendation test showcasing a 92% effectiveness rate, coupled with illustrative dialogue examples, underscore ChatDiet's strengths in explainability, personalization, and interactivity.

BACKGROUND: Alzheimers disease (AD) is a devastating neurodegenerative disorder affecting 44 million people worldwide, leading to cognitive decline, memory loss, and significant impairment in daily functioning. The recent single-cell sequencing technology has revolutionized genetic and genomic resolution by enabling scientists to explore the diversity of gene expression patterns at the finest resolution. Most existing studies have solely focused on molecular perturbations within each cell, but cells live in microenvironments rather than in isolated entities. Here, we leveraged the large-scale and publicly available single-nucleus RNA sequencing in the human prefrontal cortex to investigate cell-to-cell communication in healthy brains and their perturbations in AD. We uniformly processed the snRNA-seq with strict QCs and labeled canonical cell types consistent with the definitions from the BRAIN Initiative Cell Census Network. From ligand and receptor gene expression, we built a high-confidence cell-to-cell communication network to investigate signaling differences between AD and healthy brains. RESULTS: Specifically, we first performed broad communication pattern analyses to highlight that biologically related cell types in normal brains rely on largely overlapping signaling networks and that the AD brain exhibits the irregular inter-mixing of cell types and signaling pathways. Secondly, we performed a more focused cell-type-centric analysis and found that excitatory neurons in AD have significantly increased their communications to inhibitory neurons, while inhibitory neurons and other non-neuronal cells globally decreased theirs to all cells. Then, we delved deeper with a signaling-centric view, showing that canonical signaling pathways CSF, TGFβ, and CX3C are significantly dysregulated in their signaling to the cell type microglia/PVM and from endothelial to neuronal cells for the WNT pathway. Finally, after extracting 23 known AD risk genes, our intracellular communication analysis revealed a strong connection of extracellular ligand genes APP, APOE, and PSEN1 to intracellular AD risk genes TREM2, ABCA1, and APP in the communication from astrocytes and microglia to neurons. CONCLUSIONS: In summary, with the novel advances in single-cell sequencing technologies, we show that cellular signaling is regulated in a cell-type-specific manner and that improper regulation of extracellular signaling genes is linked to intracellular risk genes, giving the mechanistic intra- and inter-cellular picture of AD.

Exercise has beneficial effects on cognition throughout the lifespan. Here, we demonstrate that specific exercise patterns transform insufficient, subthreshold training into long-term memory in mice. Our findings reveal a potential molecular memory window such that subthreshold training within this window enables long-term memory formation. We performed RNA-seq on dorsal hippocampus and identify genes whose expression correlate with conditions in which exercise enables long-term memory formation. Among these genes we found Acvr1c, a member of the TGF ß family. We find that exercise, in any amount, alleviates epigenetic repression at the Acvr1c promoter during consolidation. Additionally, we find that ACVR1C can bidirectionally regulate synaptic plasticity and long-term memory in mice. Furthermore, Acvr1c expression is impaired in the aging human and mouse brain, as well as in the 5xFAD mouse model, and over-expression of Acvr1c enables learning and facilitates plasticity in mice. These data suggest that promoting ACVR1C may protect against cognitive impairment.

As the circadian clock regulates fundamental biological processes, disrupted clocks are often observed in patients and diseased tissues. Determining the circadian time of the patient or the tissue of focus is essential in circadian medicine and research. Here we present tauFisher, a computational pipeline that accurately predicts circadian time from a single transcriptomic sample by finding correlations between rhythmic genes within the sample. We demonstrate tauFishers performance in adding timestamps to both bulk and single-cell transcriptomic samples collected from multiple tissue types and experimental settings. Application of tauFisher at a cell-type level in a single-cell RNAseq dataset collected from mouse dermal skin implies that greater circadian phase heterogeneity may explain the dampened rhythm of collective core clock gene expression in dermal immune cells compared to dermal fibroblasts. Given its robustness and generalizability across assay platforms, experimental setups, and tissue types, as well as its potential application in single-cell RNAseq data analysis, tauFisher is a promising tool that facilitates circadian medicine and research.

Despite numerous studies in the field of dementia and Alzheimer's disease (AD), a comprehensive understanding of this devastating disease remains elusive. Bulk transcriptomics have provided insights into the underlying genetic factors at a high level. Subsequent technological advancements have focused on single-cell omics, encompassing techniques such as single-cell RNA sequencing and epigenomics, enabling the capture of RNA transcripts and chromatin states at a single cell or nucleus resolution. Furthermore, the emergence of spatial omics has allowed the study of gene responses in the vicinity of amyloid beta plaques or across various brain regions. With the vast amount of data generated, utilizing gene regulatory networks to comprehensively study this disease has become essential. This review delves into some techniques employed in the field of AD, explores the discoveries made using these techniques, and provides insights into the future of the field.

Network alignment aims to uncover topologically similar regions in the protein-protein interaction (PPI) networks of two or more species under the assumption that topologically similar regions tend to perform similar functions. Although there exist a plethora of both network alignment algorithms and measures of topological similarity, currently no gold standard exists for evaluating how well either is able to uncover functionally similar regions. Here we propose a formal, mathematically and statistically rigorous method for evaluating the statistical significance of shared GO terms in a global, 1-to-1 alignment between two PPI networks. Given an alignment in which k aligned protein pairs share a particular GO term g, we use a combinatorial argument to precisely quantify the p-value of that alignment with respect to g compared to a random alignment. The p-value of the alignment with respect to all GO terms, including their inter-relationships, is approximated using the Empirical Browns Method. We note that, just as with BLASTs p-values, this method is not designed to guide an alignment algorithm towards a solution; instead, just as with BLAST, an alignment is guided by a scoring matrix or function; the p-values herein are computed after the fact, providing independent feedback to the user on the biological quality of the alignment that was generated by optimizing the scoring function. Importantly, we demonstrate that among all GO-based measures of network alignments, ours is the only one that correlates with the precision of GO annotation predictions, paving the way for network alignment-based protein function prediction.

Most online chemical reaction databases are not publicly accessible or are fully downloadable. These databases tend to contain reactions in noncanonicalized formats and often lack comprehensive information regarding reaction pathways, intermediates, and byproducts. Within the few publicly available databases, reactions are typically stored in the form of unbalanced, overall transformations with minimal interpretability of the underlying chemistry. These limitations present significant obstacles to data-driven applications including the development of machine learning models. As an effort to overcome these challenges, we introduce PMechDB, a publicly accessible platform designed to curate, aggregate, and share polar chemical reaction data in the form of elementary reaction steps. Our initial version of PMechDB consists of over 100,000 such steps. In the PMechDB, all reactions are stored as canonicalized and balanced elementary steps, featuring accurate atom mapping and arrow-pushing mechanisms. As an online interactive database, PMechDB provides multiple interfaces that enable users to search, download, and upload chemical reactions. We anticipate that the public availability of PMechDB and its standardized data representation will prove beneficial for chemoinformatics research and education and the development of data-driven, interpretable models for predicting reactions and pathways. PMechDB platform is accessible online at https://deeprxn.ics.uci.edu/pmechdb.