Nature Biomedical Engineering

Barretts esophagus containing intestinal metaplasia predisposes to cancer, yet the majority of cases are undiagnosed. The length of a Barretts segment is a key indicator of cancer risk, but measuring it has so far relied on endoscopy, which is expensive and invasive. Cytosponge-TFF3 is a minimally-invasive test that identifies intestinal metaplasia for endoscopic confirmation. We report a machine learning technique to quantify the extent of intestinal metaplasia and predict Barretts segment length from whole-slide image tile counts automatically generated from Cytosponge-TFF3 histology slides. Utilizing data from 529 patients, our segment length prediction model achieves an average validation fold accuracy of 0.84. Applying this algorithm to an independent test set of 162 patients from a screening trial shows a precision of 0.90 for identifying short-segment disease. This advance will enable higher-risk patients to be prioritized for endoscopy while saving more than half of Cytosponge-TFF3-positive patients from endoscopy in the screening setting.

Accurate polyp segmentation from colonoscopy images is critical for colorectal cancer prevention, yet the generalization of deep learning models under domain shift remains insufficiently explored. We propose Boundary-Explicit Guided Attention U-Net (BEGA-UNet), a boundary-aware segmentation architecture that introduces explicit edge modeling as a structural inductive bias to enhance both segmentation accuracy and cross-domain robustness. The framework integrates three components: an Edge-Guided Module (EGM) with learnable Sobel-initialized operators to capture boundary cues, a Dual-Path Attention (DPA) module that processes channel and spatial attention in parallel, and a Multi-Scale Feature Aggregation (MSFA) module to encode contextual information across multiple receptive fields. Evaluated on the combined Kvasir-SEG and CVC-ClinicDB benchmarks, BEGA-UNet achieves 88.53% Dice and 82.51% IoU, outperforming representative convolutional and transformer-based baselines. More importantly, cross-dataset evaluation demonstrates strong robustness under domain shift, with BEGA-UNet retaining 83.2% of its in-distribution performance--substantially higher than U-Net (64.5%), Attention U-Net (47.5%), and TransUNet (53.1%). In a zero-shot setting on an entirely unseen dataset, the model further maintains 72.6% performance retention. Comprehensive ablation studies indicate that explicit boundary modeling plays a central role in improving generalization, while multi-scale context aggregation further stabilizes performance across domains. Feature distribution analyses support this observation by showing that edge-oriented representations exhibit markedly reduced cross-domain variability compared to appearance-driven features. Overall, BEGA-UNet provides an effective and interpretable solution for robust polyp segmentation, demonstrating that explicit boundary modeling serves as a critical inductive bias for ensuring reliability under clinical domain shifts.

Current clinical decision-making is hindered by the absence of predictive preclinical models that faithfully bridge molecular diversity to patient outcomes. Here, we apply the principle of abstraction--deriving essential features from human tissues to build next-generation new approach methodologies (NAMs) that transform patient-derived organoids (PDOs) into predictive vehicles for Crohns disease (CD). From our living biobank of adult stem cell-derived colonic PDOs, we previously defined two molecular CD subtypes: Immune-Deficient Infectious CD (IDICD) and Stress and Senescence-Induced Fibrostenotic CD (S2FCD), each defined by unique genomic, transcriptomic, and functional profiles with matched therapeutic vulnerabilities. In this study, we prospectively anchored PDO-derived molecular phenotypes to real-world clinical outcomes, revealing that S2FCD maps to baseline and progressive colonic disease activity, whereas IDICD tracks with prior ileocecal surgery, penetrating disease behavior, as well as baseline and progressive ileal disease activity. By abstracting NAMs from human tissues and cycling insights between small- n organoids and Phase 3-sized datasets, this framework recasts PDOs as dynamic, predictive platforms that capture the past, present, and future of disease behavior. Beyond oncology, this work establishes PDOs as vehicles for prospective clinical trial-like studies in inflammatory diseases and highlights colonic immune dysfunction as a potential driver of ileal CD. GRAPHICAL ABSTRACT O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=200 SRC="FIGDIR/small/25325058v3_ufig1.gif" ALT="Figure 1"> View larger version (68K): org.highwire.dtl.DTLVardef@3b0544org.highwire.dtl.DTLVardef@d6d868org.highwire.dtl.DTLVardef@119a19corg.highwire.dtl.DTLVardef@1c11082_HPS_FORMAT_FIGEXP M_FIG C_FIG

Timely detection of Barretts esophagus, the pre-malignant condition of esophageal adenocarcinoma, can improve patient survival rates. The Cytosponge-TFF3 test, a non-endoscopic minimally invasive procedure, has been used for diagnosing intestinal metaplasia in Barretts. However, it depends on pathologists assessment of two slides stained with H&E and the immunohistochemical biomarker TFF3. This resource-intensive clinical workflow limits large-scale screening in the at-risk population. Deep learning can improve screening capacity by partly automating Barretts detection, allowing pathologists to prioritize higher risk cases. We propose a deep learning approach for detecting Barretts from routinely stained H&E slides using diagnostic labels, eliminating the need for expensive localized expert annotations. We train and independently validate our approach on two clinical trial datasets, totaling 1,866 patients. We achieve 91.4% and 87.3% AUROCs on discovery and external test datasets for the H&E model, comparable to the TFF3 model. Our proposed semi-automated clinical workflow can reduce pathologists workload to 48% without sacrificing diagnostic performance.

Parkinsons disease leads to a spectrum of cardinal motor symptoms and locomotor deficits that vary in severity with the nature of daily activities and the fluctuating physiology of patients. Many of these deficits remain inadequately addressed by existing therapies that use continuous, activity-agnostic parameters. Instead, adaptive therapies embedding activity-specific parameters have the potential to better address the entire range of symptoms. Here, we expose physiological principles that enable real-time decoding of ongoing locomotor activities across motor fluctuations from the neural dynamics of the subthalamic nucleus. This decoding steered activity-dependent adaptations of deep brain stimulation therapies that improved both cardinal motor symptoms and locomotor deficits across activities of daily living. Our decoding framework provides a blueprint for next-generation neuromodulation therapies that continuously adapt parameters to the behavioral context and fluctuating physiology of each patient. One Sentence SummaryNeural decoders that leverage the physiological principles of activity-dependent encoding in the subthalamic nucleus support the implementation of adaptive deep brain stimulation therapies that alleviate locomotor deficits in people with Parkinsons disease.

Enriching diagnostically relevant cells from heterogeneous clinical samples is critical for enabling accurate detection and molecular analysis. Bioelectronic platforms offer a promising approach to this challenge by combining selective capture with real-time, label-free monitoring. We focus on Barretts oesophagus (BE), a precursor to oesophageal adenocarcinoma (EAC), where current non-endoscopic tools like the capsule-sponge yield samples dominated by background squamous cells, limiting diagnostic sensitivity. We present a bioelectronic enrichment platform that selectively captures and thermally releases columnar cells from capsule-sponge samples. The system employs a ring microelectrode array functionalised with the lectin ECA--identified here as a selective marker of columnar cells in BE--and coated with a thermo-responsive PEDOT-pNIPAAam polymer. Capture and detachment of cells were monitored using both electrochemical impedance and optical imaging, enabling real-time, label-free feedback. Applied to clinical samples, our platform enriched viable columnar cells, enhancing downstream molecular readouts. This approach integrates seamlessly with non-invasive sampling workflows and expands the utility of bioelectronic tools for early cancer diagnostics.

Fecal microbiota transplantation (FMT) is an effective therapy for recurrent Clostridioides difficile infection (rCDI) but has undefined composition and poor scalability. In vitro manufactured live biotherapeutic products (LBP) enable both scalability and defined strain composition but with higher manufacturing complexity, resulting in few LBP trials. We developed an accessible platform to produce human-grade LBPs. We provide regulatory documentation and manufacturing protocols to facilitate translating microbiome advances to human trials. With this platform, we conduct the first direct comparison of the same bacterial strains administered after in vitro manufacturing (LBP) compared to donor sourced (FMT) across two doses. In a phase 1b trial (n=18), an endoscopic dose of the 15-strain consortium MTC01 was safe with rCDI prevention eight weeks after dosing in seven out of nine LBP patients, similar to eight out of nine FMT patients. Notably, MTC01 strain engraftment was superior to FMT at higher doses.

Building brain foundation models to capture the underpinning neural dynamics of human behavior requires large functional neural datasets for training, which current implantable Brain-Computer Interfaces (iBCIs) cannot obtain due to the instability of rigid materials in the brain. How can we achieve high-density neural recordings with wide brain region access at single-neuron resolution, while maintaining long-term stability? In this study, we present a novel approach to overcome these trade-offs by introducing Fleuron, a family of ultrasoft, ultra-low-k dielectric materials compatible with thin-film scalable microfabrication techniques. We successfully integrate up to 1,024 channels within a single minimally invasive Fleuron depth electrode. The combination of the novel implant material and geometry enables single-unit level recordings for 18 months in rodent models, and achieves a large number of units detected per electrode across animals. 128-channel Fleuron probes, that cover 8x larger tissue volume than state-of-the-art polyimide counterparts, can track over 100 single-units over months. Stability in neural recordings correlates with reduced glial encapsulation compared to polyimide controls up to 9-month post-implantation. Fleuron probes are integrated with a low-power, mixed-signal ASIC to achieve over 1,000 channels electronic interfaces and can be safely implanted in depth using minimally invasive surgical techniques via a burr hole approach without requiring specialized robotics. Fleuron probes further create a unique contrast in clinical 3T MRI, allowing for post-operative position confirmation. Large-animal and ex vivo human tissue studies confirm safety and functionality in larger brains. Finally, Fleuron probes are used for the first time ever intraoperatively during planned resection surgeries, confirming in-human usability, and demonstrating the potential of the technology for clinical translation in iBCIs.

The rapid advances in multi-omics data integration technologies have opened unprecedented avenues for dissecting the mechanisms and accelerating the clinical translation of complex diseases. Nevertheless, the frequent absence of certain modalities, coupled with the inherent heterogeneity and high dimensionality of the data, severely restrict the effectiveness of integrative analysis. To address these challenges, we introduce Entropy-guided Sample-Specific Feature Selection for Incomplete Multi-Omics Learning (ESSFS-IMO), a novel framework that couples instance-wise feature selection with entropy-adaptive optimization and variational representation learning. Concretely, ESSFS-IMO leverages a Gumbel- SoftMax selector parameterized by a neural network to achieve per-sample feature selection, while an entropy-based annealing strategy adaptively controls selector sharpness. The selected features are integrated through an information-bottlenecked variational backbone with variance-weighted fusion, enabling robust classification under arbitrary missing patterns. Extensive experiments on inflammatory bowel disease (IBD) multi-omics datasets demonstrate that ESSFS-IMO consistently outperforms state-of-the-art baselines in terms of accuracy, F1 and AUC.

Movement decoding from invasive human recordings typically relies on a distributed system employing advanced machine learning algorithms programmed into an external computer for state classification. These brain-computer interfaces are limited to short-term studies in laboratory settings that may not reflect behavior and neural states in the real world. The development of implantable devices with sensing capabilities is revolutionizing the study and treatment of brain circuits. However, it is unknown whether these devices can decode natural movement state from recorded neural activity or accurately classify states in real-time using onboard algorithms. Here, using a totally implanted bidirectional neurostimulator to perform long-term, at-home recordings from the motor cortex and pallidum of four subjects with Parkinsons disease, we successfully identified highly sensitive and specific personalized signatures of gait state, as determined by wearable sensors. Additionally, we demonstrated the feasibility of using these neural biomarkers to drive adaptive stimulation with the classifier embedded onboard the neurostimulator. These findings offer a pipeline for ecologically valid movement biomarker identification that can advance therapy across a variety of diseases.

Gait dysfunction in Parkinsons disease (PD) is a major source of disability and is often resistant to traditional deep brain stimulation (DBS). Here, we report a novel neuromodulation paradigm, gait-phase-synchronized adaptive DBS (aDBS), that dynamically modulates stimulation amplitude during contralateral leg swing. In five individuals with PD, we identified personalized neural biomarkers of gait phase from cortical and pallidal field potentials and embedded them into a chronically implanted bidirectional neurostimulator. These biomarkers, derived via a data-driven search, enabled real-time detection of swing phase and sub-second modulation of stimulation amplitude. Acute in-clinic testing showed that aDBS significantly reduced gait variability and improved bilateral symmetry compared to clinically optimized continuous DBS. In a double-blinded, multi-day crossover study, gait-phase-synchronized aDBS was well-tolerated, maintained general motor symptom control, and reduced falls and improved other gait metrics. These findings establish the feasibility of biomarker-driven, movement-synchronized neuromodulation and offer a promising strategy to restore dynamic motor control in PD.

Understanding cellular architectures and their connectivity is essential for interrogating system function and dysfunction. However, we lack technologies for mapping the multi-scale details of individual cells in the human organ-scale system. To address this challenge, we developed a platform that simultaneously extracts spatial, molecular, morphological, and connectivity information of individual cells from the same human brain, by integrating novel chemical, mechanical, and computational tools. The platform includes three key tools: (i) a vibrating microtome for ultra-precision slicing of large-scale tissues without losing cellular connectivity (MEGAtome), (ii) a polymer hydrogel-based tissue processing technology for multiplexed multiscale imaging of human organ-scale tissues (mELAST), and (iii) a computational pipeline for reconstructing 3D connectivity across multiple brain slabs (UNSLICE). We demonstrated the transformative potential of our platform by analyzing human Alzheimers disease pathology at multiple scales and demonstrating scalable neural connectivity mapping in the human brain. One-Sentence SummaryWe developed an integrated, scalable platform for highly multiplexed, multi-scale phenotyping and connectivity mapping in the same human brain tissue, which incorporated novel tissue processing, labeling, imaging, and computational technologies.

Spinal cord injury (SCI) affects millions worldwide, with over half of all cases resulting in tetraplegia, where a complete injury can cause profound motor and sensory loss in all four limbs1. Here, we demonstrate an artificial double neural bypass (DNB) that integrates a bidirectional intracortical brain-computer interface with targeted spinal and brain stimulation to promote restoration of upper limb function in severe, complete paralysis. This hybrid assistive-therapeutic approach restores both hand movement and tactile sensation simultaneously via cortical mediation while promoting significant persistent sensorimotor improvements. The DNB uses a stable nested neural decoding architecture with deep reinforcement learning for fine grasping, along with patterned brain microstimulation ( cortical mirroring) and spinal cord stimulation to promote real-time and long-term functional recovery. Using the DNB, our participant with chronic C4 sensory/C5 motor complete tetraplegia regained the ability to self-feed, grasp delicate objects, and experienced persistent recovery of arm flexion and wrist tactile sensation. These findings represent a major advance in restoring meaningful function after severe, complete SCI, demonstrating that bidirectional neuroprostheses combined with targeted brain and spinal neuromodulation can drive durable sensorimotor recovery and improve independence and quality of life.

Deep learning models utilizing longitudinal healthcare data have significantly advanced epidemiological research. However, contemporary transformer-based models increasingly rely on computationally intensive pre-training steps that entail processing massive real-world datasets with cost-prohibitive hardware. We introduce the Temporal Encoder with Late Fusion (TELF), a lightweight end-to-end predictive model featuring an encoder-only architecture for processing medical codes, followed by post-encoder concatenation with demographic variables. TELF learns code embeddings on-the-fly, thereby bypassing the resource-intensive pre-training bottleneck. Furthermore, its late-fusion design preserves the integrity of the temporal attention mechanism before integrating static demographic predictors. We evaluated TELF using an administrative claims database across three distinct cohorts: pancreatic cancer (n=53,661), type 2 diabetes (n=78,756), and heart failure (n=72,540). TELF consistently outperformed traditional machine learning baselines, including XGBoost, LightGBM, and logistic regression. Specifically, TELF achieved AUCs of 0.9150, 0.8199, and 0.8721 for pancreatic cancer, type 2 diabetes, and heart failure, respectively, compared with 0.9044, 0.7908, and 0.8535 for XGBoost and 0.9014, 0.7800, and 0.8466 for logistic regression. Beyond predictive superiority, TELF's isolated temporal attention mechanism enables population-level motif mining. By extracting high-attention temporal sequences, we mapped aggregated patient journey pathways, revealing interpretable clinical trajectories preceding disease onset. Collectively, these results demonstrate that TELF provides a resource-efficient and accessible framework for advanced temporal modeling in clinical and epidemiological research.

The indicator cell assay platform (iCAP) is a novel next-generation approach for blood-based diagnostics that uses standardized cells as biosensors to amplify weak disease signals in blood. We developed an Alzheimers disease iCAP (AD-iCAP) for early detection at the mild cognitive impairment/mild dementia stages. To develop the assay, patient plasma is incubated with standardized neurons, which transduce complex circulating signals into gene-expression readouts used to train multivariate disease classifiers via machine learning. We applied systems biology analyses (e.g., GSEA, PCA, correlation/network analyses) to optimize analytical and computational parameters, and then evaluated a locked model in a study with retrospectively collected samples. Performance was AUC 0.64 (95% CI 0.51-0.78, n=82) on an independent external-validation set and AUC 0.77 (95% CI 0.57-0.96, n=23) on a blind set, supporting prospective confirmation in a larger cohort. To overcome pre-analytical noise and reduce bias in feature-selection, modeling was done using a fixed panel of 84 candidate genes chosen a priori from an external AD-iCAP dataset generated with 5XFAD mouse plasma. Despite using no AD-specific prior knowledge in this approach, the assay readout was enriched for Alzheimers-relevant pathways, including cholesterol biosynthesis, synaptic structure/neurotransmission and PIK3/AKT activation. Because the assay senses a multivalent cellular response, which is orthogonal to circulating amyloid or tau measurements, AD-iCAP may complement existing blood tests, and its multivariate readout offers a path to precision-medicine applications such as patient stratification for treatment response.

We present a Bayesian machine learning framework integrating genetic, molecular, and wearable sensor biomarkers for precision medicine in Parkinsons disease. Using PPMI (4,775 patients, 14,473 longitudinal records) and LRRK2 Consortium (627 individuals, 2,958 biological specimens), we demonstrate: (1) LRRK2 G2019S confers 1.92-fold PD risk (individual-level{chi} 2 = 36.6, p = 1.4 x 10-9; sex-adjusted OR=2.73) with carriers exhibiting 4.35-point higher motor severity (95% CI [1.95, 6.47], rank-biserial r = -0.270); (2) Wearable IMU sensors quantified Arm Swing Asymmetry (27% prevalence, n = 178) and Dual-Task Cost (14.87% degradation, t = 14.98, p < 0.001), enabling continuous cognitive-motor network monitoring; (3) Molecular markers phospho-LRRK2 (n = 884) and CSF{varepsilon} -synuclein seed amplification (n = 145) provide therapeutic monitoring and differential diagnosis; (4) Prodromal screening identified olfactory dysfunction (50.2%, n = 5, 122) and RBD (37.5%, n = 1, 548). Bayesian clustering via Evidence Lower Bound selection achieved Silhouette=0.535 with bootstrap stability (Jaccard=0.769), outperforming alternatives (0.170-0.452). Risk prediction model: AUC=0.717, calibration slope=1.197. This reproducible framework (complete code-result traceability, TRIPOD+AI compliant) enables mechanism-targeted precision medicine aligned with SDGs 3, 9, 10.

Multiplexed reprogramming of T cell specificity and function can generate powerful next-generation cellular therapies. However, current manufacturing methods produce heterogenous mixtures of partially engineered cells. Here, we develop a one-step process to enrich for unlabeled cells with knock-ins at multiple target loci using a family of repair templates named Synthetic Exon/Expression Disruptors (SEEDs). SEED engineering associates transgene integration with the disruption of a paired endogenous surface protein, allowing non-modified and partially edited cells to be immunomagnetically depleted (SEED-Selection). We design SEEDs to fully reprogram three critical loci encoding T cell specificity, co-receptor expression, and MHC expression, with up to 98% purity after selection for individual modifications and up to 90% purity for six simultaneous edits (three knock-ins and three knockouts). These methods are simple, compatible with existing clinical manufacturing workflows, and can be readily adapted to other loci to facilitate production of complex gene-edited cell therapies.

Detection and diagnosis of bloodborne pathogens are critical for patients and for preventing outbreaks, yet challenging due to these diseases nonspecific initial symptoms. We advanced CRISPR-based Combinatorial Arrayed Reactions for Multiplexed Evaluation of Nucleic acids (CARMEN) technology for simultaneous detection of pathogens on numerous samples. We developed three specialized panels that target viral hemorrhagic fevers, mosquito-borne viruses, and sexually transmitted infections, collectively identifying 23 pathogens. We used deep learning to design CARMEN assays with enhanced sensitivity and specificity, validating them and evaluating their performance on synthetic targets, spiked healthy normal serum samples, and patient samples for Neisseria gonorrhoeae in the United States and for Lassa and mpox virus in Nigeria. Our results show multiplexed CARMEN assays match or outperform individual assay RT-PCR in sensitivity, with matched specificity. These findings underscore CARMENs potential as a highly effective tool for rapid, accurate pathogen detection for clinical diagnosis and public health surveillance.

Liver transplantation (LT) is a life-saving intervention for patients with end-stage liver disease (ESLD). However, 12-20% of patients listed for LT will die on the waitlist. Modern risk scores used for transplant prioritization cannot encompass the full statistical heterogeneity of patients awaiting LT, disadvantaging women and patients with cholestatic liver disease. Our study objective was to implement more equitable LT prioritization via a more expressive class of statistical models to individualize risk prediction. To do so, we created DynaMELD, a deep machine learning-based model of waitlist prioritization. DynaMELD leverages a neural network to model complex interactions between covariates, and leverages the rate-of-change (velocity) of time-varying laboratory biomarkers to predict a more personalized risk of mortality or dropout. Our study cohort comprised 53,046 patients with ESLD listed for LT from 2016- 2023 from the U.S. Scientific Registry of Transplant Recipients. Using 90-day concordance to measure risk discrimination, DynaMELD achieves 90-day concordance 0.5% higher than MELD 3.0 (p < 0.001). Using pooled group concordance (PGCI) as a measure of fairness, DynaMELD achieves a PGCI 1.2% higher for female patients (p < 0.001), 8.3% higher for patients with primary biliary cholangitis (p < 0.001), 7.2% higher for patients with primary sclerosing cholangitis (p < 0.001), and 1.5% higher for patients with acute-on-chronic liver failure Grade 1 (p < 0.001) compared to MELD 3.0. DynaMELD reclassifies members of these sub-groups into higher risk tiers, suggesting it would improve their access to organ offers. Introspecting upon DynaMELD using the method of SHapley Additive exPlanations (SHAP) values provides an individualized degree of model interpretability. Overall, DynaMELD may provide more accurate, individualized predictions of waitlist mortality or dropout to reduce inequities and fairly prioritize patients for liver transplant.

We developed an integrated platform combining high-throughput automated biofabrication, systematic patient-derived tissue experiments, and specialized artificial intelligence to enable patient-specific computational "digital twins" for thrombosis prediction. Our automated manufacturing platform fabricates 80 fully assembled, patient-specific vessel-on-chips within 10 hours from clinical imaging--a [~]100-fold improvement over manual methods--achieving sub-micron precision through novel two-stage pneumatic motion control and integrated optical feedback. Using these chips, we systematically captured thrombosis across 491 high-fidelity videos spanning 6 patient-derived vascular geometries, 5 distinct anatomical injury sites, and 14 anticoagulant/antiplatelet interventions, establishing a "physical twin" experimental corpus. We trained CLoT (Cascade Learner of Thrombosis), a conditional video diffusion model efficiently adapted via lightweight Low-Rank Adaptation (LoRA) to generate realistic thrombosis videos conditioned on patient-specific geometry, injury location, and drug treatment. Rigorous benchmarking against state-of-the-art commercial models (Sora, Wan, Kling, Seedance, Hailuo, Hunyuan) reveals CLoT achieves 7.38-fold superior temporal biological consistency and 5.3-fold higher spatial morphological fidelity. Prospective validation on unseen patients demonstrates >90% temporal accuracy. This integrated paradigm--combining automated fabrication with domain-specialized generative AI--establishes proof-of-concept for personalized medicine enabled by digital twins trained on human-derived vascular anatomy, enabling pre-treatment antithrombotic evaluation while providing a replicable template for translating tissue engineering into clinical practice.