Objectives:We evaluated the data requirement for modern AI tools to outperform simpler models in predicting short-term mortality in over 500 000 patients with hemodialysis-dependent kidney failure. Materials and Methods:We compared logistic regression, boosting, and transformers using increasingly complex feature sets (from last-visit data to full trajectories). Performance was measured using the area under the ROC curve (AUC-ROC) and the Precision-Recall curve (AUC-PR) across training data sizes ranging from 500 to 490 197 samples. Results:Using features with temporal information is beneficial across all models. On the full dataset, Transformers (AUC-ROC = 0.8568) and boosting (AUC-ROC = 0.8598) perform similarly. Discussion:Transformers require large datasets to outperform simpler models like boosting, limiting their usefulness in smaller datasets, even on datasets as big as 500K. Conclusion:Modern AI tools require substantial data to justify their computational cost over simpler approaches. However, a more complex feature set seems to be beneficial across all models.

Oscillatory neural signals such as electroencephalography (EEG) and local field potentials (LFPs) show phase relationships that coordinate communication across brain regions. Modern recordings capture hundreds of channels across many frequency bins, yet standard phase analyses are restricted to only a few variables. The Torus Graph (TG) model, an exponential-family distribution over phases whose univariate and pairwise potentials generalize von Mises distributions, infers principled structure among oscillations but models only static, undirected dependencies and is limited to ∼ 100 variables because its score matching inference scales as 𝒪(d^6). We introduce a stochastic score matching procedure that reduces the per-iteration cost to 𝒪(d^2), enabling inference on datasets with thousands of variables. This scalable foundation supports analyses of 1,860 frequency-phase features from multi-electrode LFPs and enables two extensions previously inaccessible to TGs or classical circular statistics: (i) a TG Hidden Markov Model capturing state-dependent phase-coupling changes (e.g., spindle-related states during sleep) and (ii) an autoregressive TG inferring directional interactions via transfer-entropy estimation. Applied to LFP recordings, these models reveal state-dependent phase-interaction patterns between wakefulness and NREM sleep. Together, they enable systematic, large-scale mapping of dynamic and directional phase relationships across brain and cognitive states.

Representing the past in a compressed, efficient, and informative manner is a central problem for systems trained on sequential data. The framework, originally proposed by Gu & Dao et al., provides a principled approach to sequential compression by projecting signals onto orthogonal polynomial (OP) bases via structured linear ordinary differential equations. Subsequent works have embedded these dynamics in state space models (SSMs), where HiPPO structure serves as an initialization. Nonlinear successors of these SSM methods such as Mamba are state of the art for many tasks with long-range dependencies, but the mechanisms by which they represent and prioritize history remain largely implicit. In this work, we revisit the HiPPO framework with the goal of making these mechanisms explicit. We show how polynomial representations of history can be extended to support capabilities of modern SSMs such as adaptive allocation of memory, and input-dependent state updates, and associative memory, while retaining direct interpretability in the OP basis. We introduce a unified framework comprising five such extensions, which we collectively refer to as a "HiPPO zoo." Each extension exposes a specific modeling capability as an explicit modification of the underlying measure or dynamics governing the polynomial coefficients, rather than as an opaque learned transformation. The resulting models adapt their memory online and train in streaming settings with efficient updates. We illustrate the behaviors and advantages of these extensions through a range of synthetic sequence modeling tasks, highlighting how explicit polynomial memories can recover and clarify mechanisms implicit in SSMs.

Abstract Background Integrating multimodal data into medical artificial intelligence (AI) tools and evaluating whether they outperform human experts remains a critical challenge. Epilepsy surgery offers a unique paradigm for this evaluation, as it provides an expert-independent measure (Engel score) of post-surgical outcome. Currently, evaluation for epilepsy surgery relies on the visual interpretation and human synthesis of multimodal data. While clinical evaluations are individualized and account for complex anatomical variability, integrating these diverse, high-dimensional modalities to generate a probability of surgical success remains challenging. Here, we leverage this objective outcome score to investigate the feasibility of a data-driven, phenotype-based model against the current clinical gold standard. Methods The evaluation was performed on an epilepsy-type controlled cohort of 57 patients from six tertiary epilepsy surgery centers who underwent resective/ablative surgery in the mesiotemporal lobe. Multimodal data, namely, patient demographics, semiology, invasive electrophysiology monitoring, and neuroimaging, were utilized. We first estimated how human experts perceive surgery success. Subsequently, we developed a data-driven model integrating these modalities to predict surgery outcomes. The model performance was compared to the current clinical gold standard (three independent human experts) and published outcome calculators. Finally, modality-level phenotypes were derived based on the model’s predictions. Results Predictions by human experts correlated poorly with post-surgical outcomes, and published outcome calculators did not perform better than the experts (DeLong’s p = 0.367). Our model incorporating multimodal data achieved an area under the receiver operating characteristic curve (AUROC) of 0.801. It performed statistically better than the best human expert (DeLong’s p = 0.043) and achieved a higher AUROC than the best published surgical outcome calculator (0.801 vs. 0.694). Conclusions We demonstrated the proof-of-concept that data-driven multimodal phenotypes can inform personalized surgery planning in epilepsy. Furthermore, we provide a framework for integrating multimodal data and benchmarking medical AI performance against human experts.

Crowdsourced air temperature data from networks like Weather Underground offer dense spatial coverage and are increasingly used to study the canopy urban heat island (CUHI) effect. However, these observations are noisy: siting conditions, environmental interference, and sensor failures introduce spatially and temporally varying bias. This complicates interpolation, limiting our ability to estimate neighborhood-level air temperature. While interpolation techniques such as kriging account for uncertainty, they do so under the assumption of homoscedasticity. Moreover, they struggle to scale beyond a few thousand observations, limiting their utility on crowdsourced data. To overcome these limitations, we develop a sparse variational Gaussian process model that accounts for heteroscedasticity, allowing us to efficiently interpolate air temperature fields with calibrated uncertainty quantification. To test our approach, we apply our model to six years of hourly data across Durham County, North Carolina, and compare predictions at held-out sensor locations with linearly-interpolated ERA5-Land. Our method improves estimates at held-out locations (MAE=0.57 degrees C versus ERA5-Land MAE=3.20 degrees C) and enables high-resolution analysis of CUHI patterns over space and time. We illustrate this by visualizing (1) how CUHI patterns vary with synoptic conditions, (2) differential impacts on heating and cooling demand, and (3) annual hours exceeding 35 degrees C by neighborhood. Our method provides a scalable and statistically rigorous framework for transforming crowdsourced climate data into a gridded reanalysis product. Using this product, we can better quantify urban heat exposure and its impact on health and energy.

Air pollution is a considerable global public health threat, requiring efficient monitoring and forecasting to guide decision-making. This study introduces a cascaded model of enhanced Random Forest with Convolutional Neural Network (RF-CNN) that predicts spatiotemporal fluctuations in PM2.5 concentrations throughout Sri Lanka. The K-Nearest Neighbors method is employed to impute missing data, and the model utilizes data from 24 low-cost PM2.5 sensors that are distributed throughout the country. The Convolutional Neural Network (CNN) derives spatial features from four-band PlanetScope satellite images (3m/pixel resolution, 1km2 spatial coverage), while the Random Forest (RF) component models the relationship between PM2.5 levels and four meteorological parameters. These features, combined with meteorological, spatial, and temporal inputs, produce the final forecasting results. The dataset comprises 1934 satellite images that were collected between December 2022 and February 2024, with an average PM2.5 concentration of approximately 15 mu g/m3. The RF-CNN model exhibited robust performance metrics across a variety of climate zones, including a normalized root mean square error of approximately 32.4 %, a mean absolute percentage error of approximately 25.7 %, a normalized mean absolute error of approximately 22.8 %, a Spearman r of 0.871, and a Pearson r of 0.873. Two metrics: Input Data Quality Score (IDQS) and Testing Data Quality Score (TDQS) were implemented to evaluate the effects of imputation. Performance was minimally impacted by imputation within acceptable ranges, while exceeding limits resulted in increased uncertainty. This research emphasizes the efficacy of the RF-CNN approach, which integrates satellite imagery and low-cost sensor data, as a scalable solution for predicting spatiotemporal PM2.5 variations. It provides valuable insights for regions that lack extensive monitoring.

The field of hypothesis generation promises to reduce costs in neuroscience by narrowing the range of interventional studies needed to study various phenomena. Existing machine learning methods can generate scientific hypotheses from complex datasets, but many approaches assume causal relationships are static over time, limiting their applicability to systems with dynamic, state-dependent behavior, such as the brain. While some techniques attempt dynamic causal discovery through factor models, they often restrict relationships to linear patterns or impose other simplifying assumptions. We propose a novel method that models dynamic graphs as a conditionally weighted superposition of static graphs, where each static graph can capture nonlinear relationships. This approach enables the detection of complex, time-varying interactions between variables beyond linear limitations. Our method improves f1-scores of predicted dynamic causal patterns by roughly 22-28% on average over baselines in some of our experiments, with some improvements reaching well over 60%. A case study on real brain data demonstrates our method's ability to uncover relationships linked to specific behavioral states, offering valuable insights into neural dynamics.

Accurate and scalable quantification of animal pose and appearance is crucial for studying behavior. Current 3D pose estimation techniques, such as keypoint- and mesh-based techniques, often face challenges including limited representational detail, labor-intensive annotation requirements, and expensive per-frame optimization. These limitations hinder the study of subtle movements and can make large-scale analyses impractical. We propose Pose Splatter, a novel framework leveraging shape carving and 3D Gaussian splatting to model the complete pose and appearance of laboratory animals without prior knowledge of animal geometry, per-frame optimization, or manual annotations. We also propose a rotation-invariant visual embedding technique for encoding pose and appearance, designed to be a plug-in replacement for 3D keypoint data in downstream behavioral analyses. Experiments on datasets of mice, rats, and zebra finches show Pose Splatter learns accurate 3D animal geometries. Notably, Pose Splatter represents subtle variations in pose, provides better low-dimensional pose embeddings over state-of-the-art as evaluated by humans, and generalizes to unseen data. By eliminating annotation and per-frame optimization bottlenecks, Pose Splatter enables analysis of large-scale, longitudinal behavior needed to map genotype, neural activity, and behavior at high resolutions.

Large language models (LLMs) have recently demonstrated impressive capabilities across a range of reasoning and generation tasks. However, research studies have shown that LLMs lack the ability to identify causal relationships, a fundamental cornerstone of human intelligence. We first conduct an exploratory investigation of LLMs' behavior when asked to perform a causal-discovery task and find that they mostly rely on the semantic meaning of variable names, ignoring the observation data. This is unsurprising, given that LLMs were never trained to process structural datasets. To first tackle this challenge, we prompt the LLMs with the outputs of established causal discovery algorithms designed for observational datasets. These algorithm outputs effectively serve as the sufficient statistics of the observation data. However, quite surprisingly, we find that prompting the LLMs with these sufficient statistics decreases the LLMs' performance in causal discovery. To address this current limitation, we propose CARE, a framework that enhances LLMs' causal-reasoning ability by teaching them to effectively utilize the outputs of established causal-discovery algorithms through supervised fine-tuning. Experimental results show that a finetuned Qwen2.5-1.5B model produced by CARE significantly outperforms both traditional causal-discovery algorithms and state-of-the-art LLMs with over a thousand times more parameters, demonstrating effective utilization of its own knowledge and the external algorithmic clues.

Emotional dysregulation is highly comorbid with sleep disturbances. Sleep is composed of unique physiological states that are reflected by conserved brain oscillations. Though the role of these state-dependent oscillations in cognitive function has been well established, less is known regarding the nature of state-dependent oscillations across brain regions that strongly contribute to emotional function. To characterize these dynamics, we recorded local field potentials simultaneously from multiple cortical and subcortical regions implicated in sleep and emotion regulation and characterized widespread patterns of spectral power and synchrony between brain regions during sleep-wake states in male and female mice. First, we showed that single brain regions encode sleep state, albeit to various degrees of accuracy. We then identified network-based classifiers of sleep based on the combination of features from all recorded brain regions. Spectral power and synchrony from brain networks allowed for automatic, accurate, and rapid discrimination of wake, non-REM sleep (NREM), and rapid eye movement (REM) sleep. When we examined the impact of commonly prescribed sleep-promoting medications on neural dynamics across these regions, we found disparate alterations to both cortical and subcortical activity across all three states. Finally, we found that a stress manipulation that disrupts circadian rhythm in male mice increased sleep fragmentation without altering the underlying average brain dynamics across sleep-wake states. Thus, we characterized state-dependent brain dynamics across regions canonically associated with emotions.

Estimating counterfactual outcomes from time-series observations is crucial for effective decision-making, e.g. when to administer a life-saving treatment, yet remains significantly challenging because (i) the counterfactual trajectory is never observed and (ii) confounders evolve with time and distort estimation at every step. To address these challenges, we propose a novel framework that synergistically integrates two complementary approaches: Sub-treatment Group Alignment (SGA) and Random Temporal Masking (RTM). Instead of the coarse practice of aligning marginal distributions of the treatments in latent space, SGA uses iterative treatment-agnostic clustering to identify fine-grained sub-treatment groups. Aligning these fine-grained groups achieves improved distributional matching, thus leading to more effective deconfounding. We theoretically demonstrate that SGA optimizes a tighter upper bound on counterfactual risk and empirically verify its deconfounding efficacy. RTM promotes temporal generalization by randomly replacing input covariates with Gaussian noises during training. This encourages the model to rely less on potentially noisy or spuriously correlated covariates at the current step and more on stable historical patterns, thereby improving its ability to generalize across time and better preserve underlying causal relationships. Our experiments demonstrate that while applying SGA and RTM individually improves counterfactual outcome estimation, their synergistic combination consistently achieves state-of-the-art performance. This success comes from their distinct yet complementary roles: RTM enhances temporal generalization and robustness across time steps, while SGA improves deconfounding at each specific time point.

The exponentially increasing high-throughput single-cell RNA sequencing (scRNA-seq) data enables the further elucidation of more granular gene-cell expression patterns across diverse species, making the development of efficient cell type annotation methods an ever more pressing necessity. Although numerous high-performing annotation methods have been introduced, they continue to grapple with challenges such as dropout events or high-dimensional feature redundancy with more than 20000 genes, and are also constrained by specific limitations that may introduce manual biases. To address these challenges, we developed a deep end-to-end model (scSFUT) which can flexibly annotate scalable single-cell datasets in a purely data-driven manner based on an accuracy bias-free attention mechanism utilizing full-length gene expression. Specifically, scSFUT first performs tokenization of the gene expression vector and leverages 1D-convolution to integrate comprehensive intra-and inter-token gene pathway information. In addition, coupled with a self-supervised masking reconstruction strategy and the cell annotation task, scSFUT enables the shared encoder to obtain representative latent features at the global cell level via joint optimization with two corresponding losses. With rigorous evaluations across 5 real datasets from different species, scSFUT demonstrates competitive performance and broader applicability compared to the state-of-the-art methods.

Crowdsourced air temperature data from networks like Weather Underground offer dense spatial coverage and are increasingly used to study the canopy urban heat island (CUHI) effect. However, these observations are noisy: siting conditions, environmental interference, and sensor failures introduce spatially and temporally varying bias. This complicates interpolation, limiting our ability to estimate neighborhood-level air temperature. While interpolation techniques such as kriging account for uncertainty, they do so under the assumption of homoscedasticity. Moreover, they struggle to scale beyond a few thousand observations, which limits their utility on crowdsourced data. To overcome these limitations, we develop a sparse variational Gaussian process model that accounts for heteroscedasticity, allowing us to efficiently produce interpolated air temperature fields with calibrated uncertainty estimates. To test our approach, we apply our model to six years of hourly data across Durham County, North Carolina. This area includes a medium-sized city, so we expect our model to generalize to similarly sized regions with sufficient sensor coverage. Compared to ERA5-Land, it improves estimates of canopy temperature (MAE=0.57°C versus ERA5-Land MAE=3.20°C on held-out locations) and enables high-resolution analysis of CUHI patterns over space and time. We illustrate this by visualizing (1) how CUHI patterns vary with synoptic conditions, (2) differential impacts on heating and cooling demand, and (3) annual hours exceeding 35°C by neighborhood. Our method provides a scalable and statistically rigorous framework for transforming crowdsourced climate data into a gridded reanalysis product. Using these data, we can better quantify urban heat exposure and its impact on health and energy.

Chronic exposure to Cyanobacterial harmful algal blooms (CyanoHABs) and their cyanotoxins is an emerging global health threat. Acute exposure can cause severe neurological, gastrointestinal and respiratory illness in people and animals. Chronic exposure is associated with increased risk of liver disease, neurodegenerative diseases including amyotrophic lateral sclerosis (ALS), and cancer, although molecular mechanisms are unknown. The frequency, duration, and intensity of CyanoHABs across the globe has risen due to increased precipitation, excess nutrient enrichment, and warming temperatures, and recurrent CyanoHABs have been an ongoing concern in the Albemarle Sound region of North Carolina. To support early detection and enable system-wide CyanoHAB detection, techniques utilizing remote sensing data are essential to complement field-based methods. However, existing remote sensing approaches were primarily developed for large open-water systems and are poorly suited for dynamic, shallow environments with large land-water boundaries such as the Albemarle Sound. In this work, we developed a deep learning model using 3-band RGB Planet Lab satellite imagery collected over six sites along the Chowan River, a major tributary of the Albemarle Sound. We first validated the model on human-labeled imagery, achieving 84.09% accuracy and a 0.92 Area Under the Curve (AUC). We further validated our model by showing that predictions correlated to in situ water samples containing high concentrations of bloom-forming cyanobacterial cells. The model identified blooms in days prior to sampling, suggesting potential use for early detection and early identification of CyanoHAB events, and demonstrated statistically significant differences between interpolated cell counts during bloom events. These results show that the high resolution CyanoHAB predictive model developed here offers an approach for proactive monitoring over large rural coastal estuarine systems where local communities are more likely to experience chronic exposure to CyanoHABs.

Electroencephalography (EEG) recordings with visual stimuli require detailed coding to determine the periods of participant's attention. Here we propose to use a supervised machine learning model and off-the-shelf video cameras only. We extract computer vision-based features such as head pose, gaze, and face landmarks from the video of the participant, and train the machine learning model (multi-layer perceptron) on an initial dataset, then adapt it with a small subset of data from a new participant. Using a sample size of 23 autistic children with and without co-occurring ADHD (attention-deficit/hyperactivity disorder) aged 49-95 months, and training on additional 2560 labeled frames (equivalent to 85.3 s of the video) of a new participant, the median area under the receiver operating characteristic curve for inattention detection was 0.989 (IQR 0.984-0.993) and the median inter-rater reliability (Cohen's kappa) with a trained human annotator was 0.888. Agreement with human annotations for nine participants was in the 0.616-0.944 range. Our results demonstrate the feasibility of automatic tools to detect inattention during EEG recordings, and its potential to reduce the subjectivity and time burden of human attention coding. The tool for model adaptation and visualization of the computer vision features is made publicly available to the research community.

Fusing abundant satellite data with sparse ground measurements constitutes a major challenge in climate modeling. To address this, we propose a strategy to augment the training dataset by introducing unlabeled satellite images paired with pseudo-labels generated through a spatial interpolation technique known as ordinary kriging, thereby making full use of the available satellite data resources. We show that the proposed data augmentation strategy helps enhance the performance of the state-of-the-art convolutional neural network-random forest (CNN-RF) model by a reasonable amount, resulting in a noteworthy improvement in spatial correlation and a reduction in prediction error.

The ability to manipulate complex systems, such as the brain, to modify specific outcomes has far-reaching implications, particularly in the treatment of psychiatric disorders. One approach to designing appropriate manipulations is to target key features of predictive models. While generative latent variable models, such as probabilistic principal component analysis (PPCA), is a powerful tool for identifying targets, they struggle incorporating information relevant to low-variance outcomes into the latent space. When stimulation targets are designed on the latent space in such a scenario, the intervention can be suboptimal with minimal efficacy. To address this problem, we develop a novel objective based on supervised variational autoencoders (SVAEs) that enforces such information is represented in the latent space. The novel objective can be used with linear models, such as PPCA, which we refer to as generative principal component regression (gPCR). We show in simulations that gPCR dramatically improves target selection in manipulation as compared to standard PCR and SVAEs. As part of these simulations, we develop a metric for detecting when relevant information is not properly incorporated into the loadings. We then show in two neural datasets related to stress and social behavior in which gPCR dramatically outperforms PCR in predictive performance and that SVAEs exhibit low incorporation of relevant information into the loadings. Overall, this work suggests that our method significantly improves target selection for manipulation using latent variable models over competitor inference schemes.

The detrimental impacts of fine particulate matter (PM2.5) on human health, climate, ecosystems, crops, and building materials are well-established. However, there remain unresolved inquiries regarding the precise location of the sources of PM2.5. This study is the first attempt to use a calibrated sensors-based ambient air quality monitoring network (SAAQM network) and regulatory government monitors to train micro-satellite images for high spatial-resolution air pollution field determination of PM2.5 in Lucknow, Uttar Pradesh, India. A hybrid approach is developed to integrate three different datasets that include microsatellite images, PM2.5 ground measurements, and supporting information (meteorological parameters and geographical coordinates), to be fed into a Random Trees-Random Forest- Convolutional Neural Network (RT-RF-CNN) joint model to estimate PM2.5 concentrations at a sub-km level. The RT-RF-CNN joint model can derive PM2.5 concentrations at a spatial resolution of 500 m with statistically significant indicators such as spatial r of 0.9, a low root-mean-square error of 26.9 mu g/m(3) and a mean absolute error of 17.2 mu g/m(3). Based on our approach, the PM2.5 prediction maps using micro-satellite images (spatial resolution of 3m/pixel) and RT-RF-CNN joint model were generated for each day throughout the study period (December 2021-December 2022). The inter-grid comparison of these maps revealed the intra-urban local hotspots and coolspots at a fine-granular level seasonally, monthly, and daily. It is observed that the monsoon season has the highest number of coolspots (67%), while winter (0.1%), post-monsoon (0.5%) and summer (11%) have fewer. It is noted that the high temporal-spatial information of PM2.5 estimates from our integrated approach is not achievable by ground-based measurements and other existing satellite-based estimates alone. The findings of this study have potential applications on a diverse array, encompassing near real-time daily PM2.5 predicted maps, specific air pollution hotspot identification, PM2.5 exposure assessment at the neighbourhood level, and integration of remote sensing-based micro-satellite images and ground-based measurements.