Nature Biomedical Engineering

Intracortical brain-computer interfaces (BCIs) can restore communication to people with vocal tract paralysis by decoding cortical activity during attempted speech into text. State-of-the-art systems pairing neural-to-phoneme decoders with phoneme-to-word language models have achieved word error rates (WERs) as low as 1%, but only after collecting thousands of sentences of training data. Shortening the data collection process would facilitate scaling this new technology by reducing the time from device implant to high-accuracy communication. Here we introduce a transformer-based decoder model trained jointly across six intracortical speech BCI participants. For every participant -- regardless of sex, disease etiology, or attempted speaking strategy -- a multi-user model decoded speech more accurately (over 50% lower relative WER on average) than models trained on individual users data. Notably, the multi-user model could be finetuned on fewer than 200 sentences from a held-out user to achieve a WER below 7%. These results reveal how to pool intracortical data across people to yield more accurate, generalizable, and rapidly-deployable decoding models.

Background and Aims: Endoscopic artificial intelligence is commonly validated on selected single images, whereas gastric cancer interpretation requires integrating whole examinations. We developed GutCore and evaluated whether whole-case endoscopic images could be used for patient-level assessment of gastric cancer depth, biomarkers, and prognosis. Methods: GutCore was pretrained on 5.6 million de-identified endoscopic images from more than ten hospitals. We compared it with five general, medical, and endoscopy-specific foundation models using open image-level datasets and an internal tertiary-center cohort of 11,035 de-identified endoscopic examinations (2019-2023): 8,049 with early or advanced gastric cancer and 2,986 with benign gastritis or intestinal metaplasia. All examination images were aggregated for patient-level assessment of cancer status, invasion depth, molecular biomarkers, and overall survival. Results: Aggregating all stored images from each examination enabled patient-level gastric cancer assessment without selecting representative frames. GutCore achieved AUCs of 0.995 for cancer detection, 0.960 for muscularis propria invasion, and 0.804 for SM2-or-deeper invasion. Prediction of tissue-defined biomarker status was strongest for Epstein-Barr virus status and MLH1 loss (AUC, 0.831 and 0.854), with lower HER2 performance (AUC, 0.673). In the held-out advanced gastric cancer test set, GutCore-derived risk groups showed marked survival separation (log-rank P < .0001; high-risk vs low-risk hazard ratio, 13.18; 95% CI, 6.06-28.66), with stratification persisting within pathological stage II and III disease. External frame-level benchmarks showed strong performance for anatomical landmark recognition, disease grading, and segmentation. Conclusions: GutCore supported whole-case patient-level gastric cancer assessment using routinely stored endoscopic images. Further validation in independent clinical cohorts is needed to establish generalizability and clinical utility.

Low-intensity focused ultrasound (LIFU) is a promising neuromodulation modality, but challenges related to high response variability and the poorly understood parameter space undermine progress in clinical applications. To facilitate the development of therapeutic LIFU protocols, we developed an approach for Bayesian-enhanced adaptive control of ultrasound neuromodulation (BEACUN). BEACUN enables efficient, data-driven parameter mapping using a limited number of stimulation-response evaluations. We used functional ultrasound imaging (fUSI) to measure the neural responses to LIFU stimulation in real time, and we carried out in vivo experiments in rats to optimize and validate the performance of the BEACUN search. In live optimizations, we show that BEACUN produces more effective inhibitory LIFU neuromodulation protocols than conventional parameter exploration methods and converges to the optimal solution in 23 {+/-} 3.67 stimulation-response evaluations. Our approach realizes a platform for efficient optimization of neuromodulation parameters that could pave the way for personalized LIFU protocol development in patients.

Inflammatory bowel disease (IBD) burden is rising globally, yet only subsets of patients benefit from available therapies, underscoring the need for more precise molecular and cellular stratification. In the PREDICT study, we enrolled treatment-naive pediatric patients with IBD, alongside disorders of gut-brain interaction (DGBI) controls and healthy donors, and profiled their intestinal and blood-derived T cells using single-cell RNA sequencing (scRNA-seq). Across 107 participants, we identify a unique population of cytotoxic CD4+ T cells (CD4 CTL) enriched in the inflamed gut of patients with Crohn's disease (CD) and ulcerative colitis. CD4 CTLs are clonally expanded and express cytotoxic effector molecules and IFNG, consistent with antigen-driven activation. Cell-cell interaction analyses implicate macrophage-derived IL-27 as the top candidate for CD4 CTL differentiation, and IL-27 blockade in a mouse model limits CD4 CTL formation. Notably, elevated CD4 CTL frequencies in gut and peripheral blood at diagnosis are associated with subsequent poor outcome of anti-TNF therapy in pediatric CD. Findings in our identification cohort are validated in an independent cohort and through reanalysis of published datasets. Importantly, we designed a simple flow cytometry panel to isolate blood CD4+ CXCR6+ CD27- T cells, which displayed a CD4 CTL transcriptional phenotype. Together, our results link CD4 CTLs to anti-TNF nonresponse and support their potential as an early, blood-accessible biomarker for treatment stratification in pediatric CD.

Capturing infrequent or context dependent brain events, such as epileptic seizures and sleep abnormalities, often requires continuous monitoring over several days. It is most practical and scalable when achieved with unobtrusive, high-fidelity wireless systems that patients can use at home. Current electroencephalography systems restrict patient mobility and require continuous electrode maintenance and sub scalp solutions require surgical implantation that offer limited spatial coverage. We developed NeuroWeaves, gold polyimide microthreads thinner than a human hair that can be stitched through the epidermis using standard suture tools and connected to a lightweight wireless recorder. In preclinical models, NeuroWeaves captured whisker-evoked potentials with accuracy comparable to skull screws and matched the performance of a commercial acquisition benchmark. Semi-chronic recordings in freely moving animals remained stable for several weeks, and 30 day histology showed minimal inflammation comparable to surgical sutures. Pilot human studies reproduced posterior-dominant rhythms, photic responses, chewing artifacts, and sleep oscillations comparable to clinical electrodes. These results establish a minimally invasive, biocompatible neural interface that represents a new modality for high fidelity brain monitoring beyond conventional laboratory and clinical constraints.

The rapid expansion of chimeric antigen receptor (CAR) T cell studies has produced a fragmented evidence landscape linking publications, repository accessions, patient metadata and mechanistic observations. Here we present BioPathfinder, a multi-agent discovery engine for CAR-T research evidence construction, hypothesis generation and validation planning. Unlike existing LLM-based and agentic approaches centered on predefined CAR-T development tasks, BioPathfinder constructs a provenance-tracked resource linking scRNA-seq datasets from treated patients to their publications and uses it to generate diverse, falsifiable and dataset-aware mechanistic hypotheses prioritized for computational and experimental validation by role-specialized LLM reviewer subagents. Applied to the curated CAR-T-treated patient paper-dataset corpus, BioPathfinder nominated candidate mechanisms of CAR-T persistence, dysfunction and therapeutic resistance, including the hypothesis that genes associated with an NK-like transition program could be targeted to reduce CAR-T exhaustion and promote persistence. Patient scRNA-seq analysis showed that this NK-like transition-associated program was enriched in exhausted post-infusion CAR-T cell states. Virtual perturbation further prioritized transition-associated KLR-family receptor genes, including KLRC1, KLRD1 and KLRG1. Expert review selected KLRC1, encoding NKG2A, for experimental testing. In vitro and in vivo chronic-stimulation models showed that NKG2A marked CD8+ CAR-T cells with activated and exhaustion-associated phenotypes. NKG2A blockade improved antitumour function and persistence-associated readouts in vivo. These results show that structured clinical single-cell evidence can be transformed by domain-specialized multi-agent systems into experimentally actionable CAR-T engineering hypotheses. HighlightsO_LIBioPathfinder structures fragmented CAR-T patient evidence into a provenance-tracked paper-dataset resource. C_LIO_LIMulti-agent planning and review prioritize experimentally testable mechanisms of CAR-T dysfunction. C_LIO_LIExpert selection from BioPathfinder-nominated NK-like transition-associated genes identifies KLRC1/NKG2A blockade as a strategy to promote CAR-T persistence. C_LI

In vivo CAR-T cell therapy eliminates manufacturing complexities associated with ex vivo autologous approaches, but safety concerns have limited adoption. We developed viroVbot, a next-generation in vivo CAR-T platform, by combining computational immunogenicity prediction (CIMMEXTM) with envelope engineering. Screening 22,562 glycoprotein sequences, we identified 641 vesiculovirus homologs, from which we selected Piry virus glycoprotein (PIRYV) as the optimal candidate. PIRYV exhibited lower MHC-epitope density, reduced human seroprevalence, with decreased T cell activation compared to VSV-G. To enhance targeting specificity, we engineered receptor-binding-deficient PIRYV (ePIRYVRBD) displaying CD3/CD7 nanobodies for T cell-selective transduction. To maximize safety, we engineered CAR-TRAP producer cells to eliminate unwanted B cell transduction and incorporated machine learning-optimized T cell-specific promoters that restrict CAR activation exclusively to lymphocytes. Additional modifications suppressed hepatocyte expression and prevented phagocytic uptake. In humanized xenograft models, viroVbot3 generated potent BCMA/CD19 specific CAR-T responses against multiple myeloma and Claudin18.2-targeting gastric cancer, demonstrating sequential redosing with alternative envelopes. Critically, viroVbot3 exhibited minimal off-target organ biodistribution with CAR expression restricted to T lymphocytes. These findings establish viroVbot as a low-immunogenic platform for scalable in vivo CAR-T manufacturing with capability for sequential redosing across hematologic and solid tumors.

Adaptive deep brain stimulation (aDBS) relies on physiological biomarkers to infer motor state and guide therapeutic stimulation in Parkinson's disease. However, neural biomarkers may themselves be altered by stimulation, potentially limiting their utility for closed-loop control. We address this limitation by testing whether DBS device-embedded accelerometers can accurately track Parkinsonian motor state across stimulation conditions. We analysed over 1,900 hours of chronic recordings of subthalamic nucleus (STN), sensorimotor cortical and device-embedded accelerometry signals acquired before and during continuous STN stimulation, alongside continuous wearable assessments of bradykinesia and dyskinesia. Across stimulation conditions, accelerometry-derived features robustly tracked motor symptom severity and outperformed neural features for symptom decoding. Mechanistically, total STN beta power - a widely used biomarker for aDBS - proved less informative because it conflates periodic and aperiodic neural processes with opposing relationships to motor state. Under active stimulation, periodic beta activity showed reduced coupling to symptom severity, whereas STN aperiodic activity, cortical periodic activity and cortico-subthalamic coherence remained comparatively stable. Together, these findings demonstrate that neural and behavioural biomarkers exhibit differential robustness during deep brain stimulation and identify device-embedded accelerometry as a robust behavioural biomarker of motor state, motivating its use in next-generation adaptive DBS systems.

Many diseases arise from dysfunction of defined cell populations, yet most therapeutics distribute broadly, limiting efficacy and causing toxicity. We developed ExACT, a platform for cell-type-selective intracellular delivery that exploits membrane transporters. In vivo screening of combinatorial fluorescent small-molecule libraries in mouse brain identified chemistries whose uptake is dictated by endogenous transporter expression, yielding compounds with preferential entry into neurons, astrocytes, pericytes and endothelial cells. One series showed strong selectivity for brain and retinal endothelium, where Slco1a4 mediated uptake. This selectivity principle extended to the human orthologue SLCO1A2, highly expressed in brain endothelium and oligodendrocytes, where it mediated selective uptake in a humanized mouse model and human iPSC-derived oligodendrocytes. Ectopic expression of SLCO1A2 in neurons via gene therapy created a synthetic entry port, conferring ExACT conjugate uptake on otherwise inaccessible cells. Bifunctional compounds linking transporter-targeting motifs to antisense oligonucleotides or small-molecule drugs retained pharmacological activity while conferring transporter-dependent cell-type selectivity, illustrating how transporter diversity can be harnessed for precision pharmacotherapy.

Gastrointestinal anastomotic leakage (AL) remains a life-threatening complication following gastrointestinal surgery, where outcomes critically depend on timely diagnosis. Current diagnostic strategies rely on periodic sampling and resource-intensive analysis in centralized laboratories. Here, we present a sterilizable, time-integrating hydrogel sensor platform for continuous, infrastructure-free monitoring of the patient's postoperative drain fluid. We introduce enzyme-responsive macromolecular networks for semi-quantitative bedside assessment of leak-associated digestive enzymes. The sensors retain functionality following lyophilization and ethylene oxide sterilization, enabling long-term storage and scalable deployment around the world. In a Swiss clinical cohort of 56 patients, including 19 with gastrointestinal anastomotic leakage, the sensor detected amylase-associated leaks two days (median) prior to clinical diagnosis with a sensitivity of 78% (95% CI 55-91) and a specificity of 95% (95% CI 82-99). The prospective validation in an independent cohort of 37 patients in Colombia, including seven patients with leaks, demonstrated 100% sensitivity (95% CI 64.6-100) and a 100% negative predictive value (95% CI 87.9-100.0), with sensor activation preceding standard clinical diagnosis by a median of five days. By converting episodic biochemical measurements into continuous, cumulative visual records, this infrastructure-free material platform enables close-meshed postoperative monitoring and may facilitate earlier recognition of anastomotic leakage across diverse healthcare settings.

Despite advances in immunotherapy, most solid tumors remain resistant to treatment. Immune cell engagers redirect cytotoxic lymphocytes against cancer, but limited tumor access, immunosuppressive microenvironments and systemic immune activation limit efficacy. Here we develop live immune modulating engagers (LIME), a modular platform where non-pathogenic, tumor-tropic Escherichia coli display tandem single-chain variable fragments targeting a tumor-associated antigen and an activating receptor on T or natural killer cells. LIME bridged effector and tumor cells, induced transcriptional programs of T cell activation, metabolism and proliferation, and enhanced cytotoxicity across cancer cell lines and patient-derived organoids. In mouse models, LIME safely accumulated in tumors, outperformed tarlatamab in small cell lung cancer, and induced durable immunity in lymphoma. RAS inhibition and PD-L1 blockade enhanced LIME activity in pancreatic cancer and induced humoral responses. Multi-lineage immune modulation remained tumor-confined, without organ toxicity. These findings establish LIME as a versatile living therapeutic platform for programmable, tumor-restricted immune orchestration.

Therapeutic genome editing requires delivery of editing molecules to defined cell types, but targeting specificity and efficiency are currently limited. We hypothesized that properties inherent to immune cells, including tissue infiltration and programmed cell recognition, could be harnessed to engineer a cell-based delivery system. We show here that T cells can both produce and transfer editing machinery to target cells. In response to a programmable ligand, engineered T-lymphoid cells can transfer enzymes using complex spatiotemporal logic and deliver cargo in a cell contact-dependent or -independent manner. We demonstrate feasibility of this approach in primary human T cells, establishing a customizable genetic circuit for macromolecular delivery controlled by intercellular interactions.

Despite the significant technical advancement in spatial transcriptomics, its clinical usage is largely untapped. Here, we develop an integrated system, ENDO-Genome, for minimally invasive in-body transcript sampling to facilitate live spatial transcriptomic analysis of human internal organs. This is achieved by integrating a nanoarrayed biochip with existing endoscope to perform pressure-sensor-calibrated "Touch & Go" RNA extraction directly from human internal organs, including the highly vascularized liver or kidney, without the need for tissue biopsy, voiding any bleeding risks. By a demonstration using gastrointestinal endoscopy, multiplexed landscape of 55 mRNA transcripts was obtained from multiple locations of human intestinal tract via a 5-minute operation in routine examinations. Benefiting from a sequencing-free approach, each assay costs less than 10 US dollars. For the clinical study involving 15 Crohn' s disease (CD) patients, no complication case was reported out of 47 ENDO-Genome operations, showcasing the gentle deposition and excellent safety of the technique. The live spatial transcriptomics provides direct in vivo pictures of the heterogenous spatial transcriptional programs underlying CD pathological response at different intestinal locations, revealing distinct ileal phenotypes. This is manifested by unique microscale scattering of inflammation gene clusters, along with the discovery of a tissue-specific cooperative mechanisms between inflammation and RNA methylation regulations at single- or multi-cell scales.

Pancreatic ductal adenocarcinoma (PDAC) remains a leading cause of cancer-related mortality, largely due to diagnosis at advanced stages. Early detection through minimally invasive liquid biopsy holds promise for improving patient outcomes. Here, we report a blood-based liquid biopsy platform based on single-molecule enzyme activity profiling (SEAP), which detects proteoform-level alterations in circulating pancreatic enzymes at single-molecule resolution. Using a tissue-centric biomarker discovery strategy, we identified activity signatures of pancreas-specific digestive enzymes associated with PDAC. A combinatorial classifier detected PDAC (stage I-IV) with 95.5% specificity and 75.0% sensitivity (74.2% for stage I-II) across 690 blood samples collected from multiple hospitals and biobanks. Performance was further validated in an independent cohort enriched for early-stage disease, where 54.2% (13/24) of stage IA, 64.3% (9/14) of stage IB, and 21.4% (3/14) of stage 0 lesions were classified as positive. These findings support the clinical potential of SEAP for early detection of PDAC.

Hyperpolarised magnetic resonance (MR) enables real-time measurement of metabolic flux in living systems but remains difficult to deploy for cell-based studies because each hyperpolarisation event typically interrogates a single biological condition, limiting throughput, replication and longitudinal experimentation. Here we present a deployable microfluidic platform for parallel and longitudinal hyperpolarised 13C metabolic phenotyping using standard MRI instrumentation. By combining microfluidics with spatially resolved MR spectroscopic imaging, the platform converts a single hyperpolarised preparation into multiple independent metabolic measurements without dedicated radiofrequency receive arrays or specialised instrumentation. We demonstrate reproducible discrimination of metabolically active and inactive samples, resolve cell-type-specific metabolic phenotypes, quantify biochemical and pharmacological perturbations, and recover metabolic exchange kinetics from parallel samples. Beyond increasing throughput, the platform enables repeated, non-destructive metabolic interrogation of the same recirculating three-dimensional cell cultures, allowing longitudinal phenotyping of living constructs rather than endpoint comparisons of independent samples. Across this study, 186 HP-MR measurements were acquired using only 44 polarisation events, corresponding to an approximately 4-fold increase in experimental throughput, while longitudinal monitoring reduced biological sample preparation 5-fold by following the same constructs over time. By lowering the technical barrier to hyperpolarised metabolic imaging while enabling both parallel and longitudinal metabolic phenotyping, this platform provides an accessible framework for drug discovery, microphysiological systems and patient-derived models.

BackgroundThermally ablative focused ultrasound (T-FUS) offers a noninvasive, spatially precise strategy for local tumor destruction, with the added potential to remodel tumor architecture and immune dynamics in ways that influence downstream therapeutic delivery and efficacy. Despite promising preclinical and clinical findings, the T-FUS parameters that best balance tumor debulking with preservation of local biologic, e.g. immunotherapy, penetrance remain unclear. Thermal dose, defined by the relationship between tissue heating, exposure duration, and biological effect, is likely a critical determinant of this balance. Excessive thermal dose may eliminate the vascular and stromal features needed to support immunotherapy access, whereas insufficient thermal dose may fail to achieve meaningful cytoreduction. Here, we deploy multimodal PET, contrast-enhanced ultrasound, and tissue profiling to define a "Goldilocks Zone" for T-FUS that balances bulk tumor destruction with immunotherapy delivery. MethodSubtotal T-FUS was applied to 4T1 tumors using three thermal dose regimens resolved by in silico modeling. Ablation was quantified by H&E and TTC staining. Post-ablative perfusion and microvascular coverage were assessed by contrast-enhanced ultrasound and immunofluorescence, respectively. Tumor oxygenation was measured by intravenous hypoxyprobe labeling. After T-FUS, mice underwent dynamic [18F]-FDG PET and immunoPET with a model tumor-targeted antibody, [89Zr]-CD47, to relate cytoreduction to antibody penetrance. ImmunoPET findings were further evaluated by ex vivo biodistribution analysis. ResultsIn silico modeling established three T-FUS regimens that generated distinct thermal dose profiles and were deployed in vivo in a solid breast tumor model. Histopathology, perfusion imaging, and hypoxia analysis revealed dose-dependent and dose-divergent biological effects that informed a candidate Goldilocks thermal window. Low thermal dose produced measurable but limited tumor debulking, whereas high thermal dose caused disproportionate functional perfusion collapse. An intermediate thermal dose achieved robust partial ablation, broad hypoxia relief, and preservation of residual tumor physiology sufficient to support antibody access. Dynamic [18F]-FDG PET confirmed a marked reduction in metabolically active tumor burden after Goldilocks T-FUS. Serial [89Zr]-CD47 immunoPET showed that bulk antibody signal was maintained after ablation, and integration of immunoPET with matched [18F]-FDG PET revealed approximately 3-fold enrichment of antibody exposure within the residual viable tumor compartment of ablated tumors. These findings demonstrate that appropriately tuned thermal ablation can debulk tumor while preserving, and potentially concentrating, immunotherapy access within the remaining targetable tumor niche. ConclusionThis study identifies thermal dose as a critical consideration for T-FUS immunotherapy combinations and establishes a PET-informed framework for balancing cytoreduction with therapeutic delivery. Rather than functioning solely as a local debulking modality, we demonstrate that T-FUS can be tuned to yield a post-ablation tumor state that remains accessible to large biologics. These findings provide timely, translationally relevant guidance for tailoring T-FUS regimens to achieve local tumor destruction while preserving an immunotherapy-permissive niche for combination treatment.

Miniaturized implantable bioelectronics offer breakthrough potential in disease treatment, biomarker monitoring, and physiological sensing. However, wireless power transfer (WPT) remains a central limitation for millimeter-scale devices, as scaling down the receiver size rapidly decreases coupling efficiency due to tissue attenuation, low quality-factors, reduced mutual inductance, and limited tolerance to spatial and angular displacement. Here, we introduce distributed resonant coupling (DRC), a 3-coil WPT paradigm which enables the receiver (Rx) to actively participate in a strongly coupled resonance, transforming the Rx from a passive energy harvester into an active participant in a strongly coupled regime. By co-designing transmitters (Tx), resonators (Rs), and mm-scale receiver coils (Rx) as fully coupled systems, DRC exhibits simulated maximum power levels of 66%, measured received power transfer efficiencies of [~]56%, and end-to-end DC power transfer efficiencies of 42%, delivering >420 mW to loads at 1 W of transmitted power while maintaining robust performance across a range of practically relevant orientations and tissue media. Systematic theoretical and experimental efforts establish core design rules for DRC systems enabling operation without specialized tuning integrated circuits or components. To illustrate the capabilities of DRC in practical applications in vivo, we demonstrate three technologies that capture a broad application space in bioelectronics: implant localization, rapid wireless battery charging, and ultraminiaturized drug delivery devices compatible with particulate drug formulations. Taken together, these results suggest operational capabilities across a wide range of angular tolerances (up to 60{degrees}), tissue depths and dielectric and scattering media.

Treatment with chimeric antigen receptor (CAR) T cells has emerged as a promising immune therapy for relapsed and refractory hematologic malignancies. The CAR T cells are manufactured in a series of steps that involve isolating T cells from the patients leukapheresis product, genetically modifying them to express the CAR against the target antigen, and reinfusing them into the patient. Efficient T-cell enrichment from leukapheresis products is critical to the success of these therapies. Current methods for T-cell sorting on a clinical scale involve several washing steps to remove red blood cells and platelets, followed by T-cell selection and activation. These multi-step processes result in cell loss during processing and involve several handling steps. Here, we utilize fluidically assembled micromagnetic lenses to develop a high-throughput, continuous-flow microfluidic T-cell sorter, designated as the T-Chip, for sorting magnetic bead-labeled CD3+ T cells in a single step. Our approach allows direct sorting of T cells in expansion media from leukopaks without any washing steps, effectively removing 99.999% of RBCs and platelets from the leukapheresis product. A single 1-inch x 3-inch T-Chip can process leukapheresis product at a throughput of 60 mL/hr and 2.56 {+/-} 0.12 billion cells/hr. Using this optimized workflow, we demonstrate clinical-scale enrichment of highly pure CD3+ T cells (97.7 {+/-} 1.3%) with high viability (97.0 {+/-} 1.1%) and recovery (87.3 {+/-} 14.8%) in a functionally closed manner. Downstream processing of T cells isolated using the T-Chip yielded potent anti-mesothelin CAR T cells with demonstrated anti-tumor efficacy. Overall, by exploiting precisely engineered magnetic forces and laminar flow, the microfluidic T-Chip overcomes bottlenecks caused by low throughput and enables single-step large-scale T-cell purification for the rapid development of CAR T cells.

Body surface gastric mapping (BSGM) enables non-invasive measurement of gastric electrophysiology, but the signals are approximately 100 times weaker than cardiac potentials and overlap spectrally with motion artifacts, necessitating labor-intensive manual review that limits clinical scalability. We present an uncertainty-aware deep learning framework combining a signal reconstruction network with a parallel uncertainty estimation network to automate artifact correction in high-resolution BSGM. Models were trained on 2,398 multihour, 64-channel recordings from 27 international clinical sites using weak supervision, a physiology-aware loss function, and uncertainty-gated quality control. In an independent cohort of 127 patients, the system achieved relative reductions of 39% in signal reconstruction error, 9% in total data removed, and 23% in amplitude--movement correlation compared with the industry-standard Wiener filter. Improved signal fidelity altered automated clinical phenotyping in 7% of patients by recovering previously obscured gastric rhythms. Uncertainty-aware deep learning enables reliable automated artifact correction in body-surface gastric mapping, improving signal fidelity and enabling scalable clinical interpretation. The system is FDA-cleared (510(k) K252504) and deployed in clinical practice, demonstrating that data-driven artifact correction can meet regulatory requirements for medical devices and reduce dependence on specialist manual review.

RNA interference (RNAi) therapeutics enable selective silencing of disease-associated genes. Yet, their clinical application remains largely confined to the liver due to extrahepatic delivery constraints of current platforms such as GalNAc conjugates and lipid nanoparticles. Extracellular vesicles (EVs) offer an attractive alternative delivery strategy owing to their biocompatibility, ability to traverse biological barriers, and amenability to engineering. However, EV-mediated RNA delivery is limited by inefficient endogenous RNA loading and poor cytosolic release following uptake. Here, we establish a modular EV-based platform that addresses both challenges by integrating enhanced endogenous shRNA loading with fusogen-mediated cytosolic delivery. Using Argonaute 2 (AGO2)-assisted loading, we substantially increase shRNA copy numbers per vesicle (up to 3.7 copies/EV) and enable quantitative, molecule-resolved assessment of delivery potency. Engineered EVs achieve robust and reproducible shRNA-mediated gene silencing with picomolar IC50 values across multiple cell types and induce significant target knockdown in the mouse brain following intracerebral administration. Together, these findings demonstrate that coordinated engineering of shRNA loading and cytosolic release can overcome key limitations of EV-mediated small RNA delivery.