Nature Medicine

Cellular senescence is a central hallmark of aging, yet its measurement in humans remains invasive, low-throughput and tissue-specific, precluding population-scale study. We developed Methylation-associated Gene Expression (MaGE) predictors: whole-blood DNA methylation proxies of p14ARF, p16INK4A and p21CIP1 expression, plus a composite score, providing the first scalable measure of senescence-marker expression. Deployed across Generation Scotland (n=18,859), MaGE yielded 45 Bonferroni-significant associations spanning 18 incident diseases and all-cause mortality. The two strongest associations recapitulated the established tissue specificity of p21CIP1 and p16INK4A activation, with MaGE-p21 associating with incident alcoholic liver disease and MaGE-p16 with pulmonary fibrosis. In a separate study, MaGE tracked disease severity and treatment response in Crohn's disease. MaGE is a scalable, interpretable biomarker of senescence that enables its study at population scale and offers a route to patient stratification in senolytic trials.

Multimodal Parkinson disease models can improve internally while obscuring why patients differ and whether a gain is deployable. We separated clinical trajectories, dated biological sources, observation process and cohort-relative, age-weighted evidence for dopaminergic, injury/glial, cognitive co-pathology and lysosomal burden across 5,404 cutoffs in 1,058 Parkinson's Progression Markers Initiative participants. DaT-SPECT and MoCA history produced a small 12-month motor-change gain (Delta R^2=0.015; delta MAE=-0.060) that failed prespecified magnitude and strict-calendar criteria. Continuous evidence retained overlapping support but showed no detectable predictive increment beyond raw sources. Source-aware intervals achieved 90.6% simultaneous participant coverage yet failed selective prediction, yielding ABSTAIN. In 8,000 semisynthetic experiments, the gate limited null promotion to 0.35%, detected leakage and blocked 94.75% of calendar reversals; PDBP clinical transport and multimodal eligibility failed. The framework reveals state support, stale or missing evidence, and unjustified prognostic, uncertainty or transport claims.

The World Health Organization has called for malaria vaccines with > 90% efficacy against Plasmodium falciparum (Pf) infection. The live, non-attenuated PfSPZ-CVac (chloroquine) vaccine provides the strongest protection against controlled human malaria infection (CHMI). Its 100% protective efficacy at 12 weeks against genetically distant Pf stems from a highly amplified immune stimulus. After inoculation, each invading sporozoite (SPZ) replicates up to 50,000 times within a hepatocyte. Over six days of development, these parasites express thousands of distinct proteins, presenting a vast and diverse array of antigens to the host immune system. The late liver stage parasites then rupture into the bloodstream, where they are killed by the co-administered drug, chloroquine. To engineer a Pf parasite that develops to the late liver stage like PfSPZ-CVac (chloroquine), but then arrests before transitioning to the blood stage, two genes required for transition from non-pathogenic liver to pathogenic blood stages were identified and deleted, creating late-liver stage arresting, replication-competent (LARC) PfSPZ. In a first-in-human, randomized, double-blind, placebo-controlled clinical trial, we assessed PfSPZ-LARC2 Vaccine in 75 malaria-exposed Burkinabe 1- to 48-year-olds. No vaccine-strain breakthrough blood-stage infections were detected; safety and tolerability profiles were similar in vaccinees and placebo recipients; and the highest immunogenicity was observed in the youngest recipients. Based on these findings, PfSPZ-LARC2 Vaccine has progressed to Phase 2 testing with the goal of rapidly advancing this genetically attenuated whole parasite vaccine to licensure and deployment.

SARS-CoV-2 antigenic evolution continues to challenge strain-matched vaccination. For the 2026-27 season, the European Medicines Agency (EMA) identified XFG as the preferred antigen within the JN.1 family and Comirnaty XFG was authorised in the European Union in July, 2026. We studied 22 participants who received 30 microg of mRNA encoding XFG (Comirnaty XFG, BioNTech/Pfizer, Mainz, Germany) in September, 2026 and had paired serum available before vaccination and median 14 days afterwards. Median Wuhan-Hu-1 S1-specific IgG increased from 1600 BAU/mL (IQR 1118-2926) to 2914 BAU/mL (2117-4468), a 1.8-fold increase. Median RBD-binding ED50 increased from 116 (51.4-212) to 305.6 (170.8-434.6) for LP.8.1 and from 354.3 (129.2-558.2) to 832.9 (429-1175) for NB.1.8.1, corresponding to 2.6-fold and 2.4-fold increases, respectively (all p<0.0001). XFG-adapted vaccination increased neutralising activity against each tested variant. The GMT against XFG increased from 180 before vaccination to 894 afterwards (5.0-fold), and that against RV.1.3 increased from 121 to 599 (4.9-fold). GMTs increased from 117 to 386 for PQ.16.1.1 (3.3-fold), from 101 to 316 for XFG.1.1 (3.1-fold), from 629 to 1772 for LP.8.1 (2.8-fold), from 221 to 611 for NB.1.8.1 (2.8-fold), and from 228 to 642 for PJ.2.1 (2.8-fold). Neutralisation of BA.3.2.2 increased more modestly, from GMT 78 to 126 (1.6-fold; p=0.0009). Pre- versus post-vaccination neutralisation titres differed significantly for all variants (p<0.001). Despite the broad increase, post-vaccination neutralisation remained heterogeneous. Relative to LP.8.1, GMTs were 2.0-fold lower for XFG, 2.8-fold lower for PJ.2.1, 2.9-fold lower for NB.1.8.1, 3.0-fold lower for RV.1.3, 4.6-fold lower for PQ.16.1.1, 5.6-fold lower for XFG.1.1, and 14.1-fold lower for BA.3.2.2. These findings provide early human immunogenicity data after XFG-adapted mRNA vaccination. The largest relative increases were observed for XFG and RV.1.3, whereas LP.8.1 remained the most strongly neutralised variant after Comirnaty XFG vaccination and BA.3.2.2 the least. This pattern suggests broad boosting of cross-reactive humoral immunity without elimination of between-variant differences. The small, extensively pre-immunised cohort, short follow-up, absence of a contemporaneous vaccine-comparator group, and use of a pseudovirus assay limit inference. Our data characterise early antibody binding and neutralisation but do not establish durability or clinical effectiveness.

Multi-agent prompting of large language models has produced contradictory diagnostic results, and it remains unclear whether any benefit comes from specialist personas or from perspective independence. We compared a single direct call, five personas in one context, and the same five roles as isolated agents integrated by a moderator, with five repeat runs per case, on 87 CPC cases, 406 MedCaseReasoning cases, and 364 emergency department encounters, under an LLM judge validated against clinicians. On the external benchmark the team beat the single call on both pre-specified recall endpoints (top-3 +3.0 points, p = 0.0079; top-5 +3.9, p = 3.8 x 10^-5;); a factorial attributes the gain to independent generation plus moderated synthesis, not the specialist roles. On real emergency presentations the benefit reversed (top-1 40.1% versus 34.3%, p < 0.0001), carried by the specialist role lists and surviving added objective results. Deployment should key on the question and the input at hand.

Mammography interpretation requires tying each finding to supporting evidence and withholding judgment when that evidence is unavailable. Vision-language models are increasingly used for image-based medical question answering, yet existing benchmarks suggest that they struggle to jointly support accurate answers, evidence grounding, and appropriate abstention. We propose CARE-MVLM, built on Qwen2.5-VL-7B-Instruct, to address this gap through separate modules for answer prediction, answer-conditioned evidence generation, and answerability selection. On 7,358 triplets curated from CBIS-DDSM, MIAS, and VinDr-Mammo, CARE-MVLM substantially improves lesion-sensitive selective behavior over same VLM backbone baselines and exhibits markedly stronger lesion-sensitive abstention than GPT-5.6 and Gemini-3.5-flash while attaining the highest joint answer, grounding, and abstention reliability in the comparison.

Background Exposure-history data collected during epidemics or pandemics within digital contact-tracing applications hold immense potential for rapid, scalable epidemiological inference. The sensitivity of these data makes privacy-preserving designs essential to secure public trust and app uptake. However, during the COVID-19 pandemic, such approaches were not yet available at the scale or maturity required for deployment. We present a Bayesian decentralised framework to infer transmission dynamics from exposure-history data while protecting user privacy. Methods Our approach is based on iterative communication between user devices and a central server, computing log-likelihood gradients on device from exposure-history data and centrally updating parameter values using only aggregated gradients. Applying this framework to synthetic exposure-history data from a hypothetical respiratory pathogen epidemic simulated on dynamic contact networks, we assessed its ability to estimate the underlying transmission parameters, perform infector attribution, and reconstruct serial-interval (SI) distributions. We systematically explored how inference was affected by incomplete exposure-history data arising from partial app adoption, case under-ascertainment, and incomplete exposure logging, and by transmission scenarios. Findings Our decentralised framework accurately recovered transmission parameters, closely matching estimates derived from centralised data with no systematic bias in posterior means. Parameter recovery remained robust when hazards from unobserved exposures were known as a daily average across the population. Although incomplete exposure-history data and specific transmission scenarios introduced compensatory parameter biases, the inferred host infectiousness profile retained sufficient accuracy, and the weekly SI distributions reconstructed from infector attributions closely resembled the ground-truth distributions. Interpretation Our findings highlight that, beyond its primary role as a control measure, digital contact tracing could help characterise transmission dynamics almost in real time and inform public health interventions while preserving user privacy. Funding Oxford Martin School Programme on Digital Pandemic Preparedness; National Institute for Health and Care Research (NIHR) Health Protection Research Unit in Emerging and Zoonotic Infections.

Survival prediction frequently exhibit unstable and irreproducible performance, limiting their clinical applicability. Rather than proposing a new predictive architecture, this work presents a diagnostic investigation into the sources of this instability. Using a transformer-based deep fusion model combining RNA sequencing and histopathology data from the TCGA-BRCA cohort as a diagnostic probe, we demonstrate that global survival prediction is highly unstable across data splits. We then analyze the learned multimodal embeddings in an unsupervised manner, revealing stable latent patient subgroups with significantly divergent survival outcomes. These subgroups are not explained by canonical receptor status or immune infiltration, but instead align with tumor-intrinsic proliferative activity. Conditioning survival modeling on the largest discovered subgroup recovers strong and stable prognostic performance, substantially outperforming unimodal baselines. These findings indicate that survival predictability is conditional on latent heterogeneity and suggest that heterogeneity-aware modeling strategies are essential for robust multimodal prognostication. This work establishes a methodological framework for interpreting instability in multimodal survival learning.

Blood metabolomics capture variation in biological aging, but the circulating proteins that underlie or accompany metabolomic aging and its links to disease and mortality remain underexplored. Here, we integrated metabolomics, proteomics and genetics from UK Biobank to characterize the protein architecture of metabolomic aging. Building on our previous mortality-trained score (MetScore1), we developed MetScore2 from 325 NMR biomarkers in 400,594 participants. Higher MetScore2 values marked earlier onset of age-related diseases, and faster score increases predicted higher mortality. Proteome-wide association study identified MetScore2-associated plasma proteins. Their signature was broadly similar in both sexes but showed undulating changes across mid-to-late adulthood, with peaks at 50, 57 and 65 years. By integrating GWAS, Mendelian randomization and colocalization, we identified IL1RN, ARG1, NCAN, F11R, SERPINF2 and NBL1 as putative causal proteins, each with distinct disease associations. Our findings delineate the proteomic landscape of metabolomic aging and prioritize candidate causal proteins for mechanistic and translational follow-up.

Biological aging clocks typically summarize complex biological profiles as a single age estimate, yet how uncertain these estimates are and whether this uncertainty represents a distinct, biologically meaningful dimension of human aging remain unclear. Here we develop a probabilistic framework to derive biological age gap (BAG) and biological age uncertainty (BAU) from the mean and width of predicted-age distributions in more than 450,000 UK Biobank participants, evaluate their prospective associations with aging-related outcomes, and characterize the molecular and genetic foundations of BAU through cross-omic profiling, genome-wide association studies and tissue and functional mapping. BAU was weakly correlated with BAG and increased with age-signal disagreement across individual molecules and discrete organs, linking uncertainty to reduced coherence of aging biological systems. Higher BAU was associated with mortality, loss of healthspan and multimorbidity independently of BAG, with mortality hazard ratios of 1.24 (95% CI, 1.23-1.26) and 1.40 (95% CI, 1.37-1.43) per standard deviation for NMR-derived and proteomic-derived BAU, respectively. Longitudinal increases in NMR-derived BAU were associated with subsequent health decline, and mortality associations were replicated using routine-blood-derived BAU in three independent external cohorts. Cross-omic analyses linked BAU to inflammation, metabolic regulation and tissue remodelling. Genetic analyses revealed overlapping but distinct architectures of BAU and BAG, with tissue and functional mapping further implicating immune and lipid-metabolic processes. This study establishes a framework for characterizing biological age through both its mean deviation and uncertainty, defining BAU as a new dimension of human aging that links the coherence of aging signals to future health vulnerability.

Sepsis causes approximately 11 million deaths globally each year. Whole-genome sequencing (WGS) of founder populations offers a unique opportunity for discovery of genes related to susceptibility of sepsis. We performed GWAS and EXWAS on WGS data for n=210 Arab ICU patients (n=68 sepsis cases, n=142 controls, other ICU-related disorders) and prioritized n=40 associated deleterious founder variants in n=31 genes (REGENIE Firth regression p < 0.05; CADD score > 20; enriched AAF vs AllofUs/gnomAD), including a novel start-loss variant in UBE2L3 (c.2T>G; p.Met1?; CADD = 25.6; SIFT = 0.001) in sepsis cases (3/68 vs 0/142 controls; Fisher's exact P = 0.026). In All of Us (AoU), UBE2L3 missense burden was associated with bacterial infection (n=26,939 cases; SKAT-O P = 2.38x10-3; OR=23.10; 95% CI = [0.63-845.27]). LLM AI models fine-tuned for drug development scored UBE2L3 a 7.8 of 10 as a drug target for sepsis, consistent with mouse UBE2L3 KO showing increased IL-1Beta mediated inflammation in sepsis.

Background Emerging evidence has linked frailty to individual chronic diseases, but whether it predicts a broad disease spectrum and accelerates multimorbidity accumulation over time remains unknown. This study aimed to systematically evaluate the associations of physical frailty with the burden of incident disease across a wide range of chronic conditions, and with the long-term rates of multimorbidity accumulation. Methods In this prospective cohort study of 466,864 UK Biobank participants , frailty status was assessed using the Fried frailty phenotype and categorized as robust, pre-frailty, or frailty. Chronic diseases were defined based on International Classification of Diseases (ICD)-10 codes, and disease patterns were identified using network analysis. Cox proportional hazards models were used to examine the associations of frailty status with incident chronic diseases, and linear mixed models were used to assess the annual rate of chronic disease accumulation. Results During a median follow-up of 13.8 years (IQR: 13.1-14.5), 536 incident chronic diseases across multiple systems were recorded. Physical frailty was associated with a significantly increased risk of 458 incident diseases (Bonferroni-corrected P < 0.05/536). In the overall cohort, compared with robust participants, those with pre-frailty ({beta} = 0.266, 95% confidence interval [CI]: 0.264-0.267) and frailty ({beta} = 0.941, 95% CI: 0.936-0.946) exhibited significantly faster rates of multimorbidity accumulation. This accelerated multimorbidity accumulation was consistently observed across strata of baseline disease burden, including participant with 0 ({beta} = 0.644, 95% CI: 0.635-0.653), 1 ({beta} =0.748, 95% CI: 0.727-0.768), [&ge;] 2 ({beta} =0.940, 95% CI: 0.933-0.947) chronic conditions at baseline. Similar associations were also found across distinct disease patterns, including neurodegenerative, metabolic, and psychiatric patterns. Conclusions Physical frailty predicts a broad range of incident diseases and is associated with accelerated multimorbidity accumulation. Incorporating frailty assessment into prevention strategies may help mitigate the growing burden of multimorbidity.

Foundation models for colour fundus photography (CFP) are candidate visual encoders for multimodal medical AI, but existing models are large and are assessed under fine-tuning rather than the frozen-encoder conditions such systems impose. We present RFM (Retinal Foundation Model), a 64-million-parameter Vision Transformer initialised from DINOv3-Base and domain-adapted with the LeJEPA joint-embedding predictive self-supervised objective with domain-specific pretext tasks, on 4.3 million CFPs from the North East London Diabetic Eye Screening Programme. A single linear head per task on one frozen backbone measures representation generalisability, not fine-tuning capacity. Across six external datasets spanning diabetic retinopathy (DR) grading, glaucoma detection, ten UK Biobank biomarker phenotypes and 3-year incidence of myocardial infarction, ischaemic stroke and Parkinson's disease, RFM attains the lowest mean absolute error of any frozen model on all seven regression phenotypes and a higher area under the receiver operating characteristic curve (AUROC) than any RETFound variant on all eight matched classification benchmarks. It also attains the highest AUROC of any model evaluated on four of the five ophthalmic benchmarks, with 4.8x fewer parameters than RETFound-MAE and RETFound-DINOv2. Applied to UK Biobank without retraining, RFM exceeds the best published retinal foundation-model AUROC for 3-year myocardial infarction and ischaemic stroke; the Parkinson's endpoint (11 cases) is inconclusive. DINOv3-Base without retinal pre-training, read the same way, stays competitive on DR grading but falls clearly behind on glaucoma, and the feature-extraction convention alone is worth up to 0.154 AUROC. One lightweight backbone yields generalised representations serving ophthalmic grading, biomarker regression and incidence prediction.

Environmental exposures shape disease risk, but the generalizability of circulating protein-based readouts of exposures across diverse populations remains unknown. Using plasma proteomic profiles of 5,416 proteins in 10,043 participants from the All of Us Research Program, we evaluated associations with and composite scores of 100+ exposures, spanning adiposity, air pollution, ambient temperature, and substance use. We find that proteomic exposure signatures are largely consistent across populations and socioeconomic strata, with <1% of proteins showing significant cross-population heterogeneity, which included BMI as an exception, with leptin being the most heterogeneous association. Population-specific leptin trajectories adiposity, together with residual leptin-type 2 diabetes associations after BMI adjustment, revealed sex-specific metabolic risk not captured by BMI alone. Aggregated protein scores were largely exposure-specific, but substance-use scores showed cross-prediction and SES scores were broadly associated. PM2.5 protein scores captured within-region exposure variation and outperformed geographic and socioeconomic correlates, but spatial confounders remain embedded within protein-environment associations. Scores trained in UK Biobank transferred to All of Us for biologically proximal exposures, supporting cross-biobank portability. These findings establish plasma proteomics as a scalable framework for exposure measurement across diverse populations, while underscoring that responsible applications require careful consideration of biological context, social patterning, and cohort structure.

The causes of most cancers remain unknown because the biological pathways underlying cancer development are poorly understood. Large-scale plasma proteomics offers a powerful strategy to uncover these pathways, identify biomarkers for cancer early detection, and discover novel therapeutic targets but has rarely been applied in prospective cohorts. Here, we profiled 6,402 plasma proteins using the SomaScanTM 7K Assay in baseline samples from 9,828 participants in the European Prospective Investigation into Cancer and Nutrition (EPIC), including 6,073 incident cancer cases within a case-cohort design. We identified 1,732 protein-cancer associations (FDR P < 0.05) across 25 cancer endpoints in multivariable-adjusted models, including 60 proteins associated with overall cancer risk. Confirmation analyses in the Atherosclerosis Risk in Communities (ARIC) Study cohort showed directional concordance for 78% of associations and statistical replication of 49% of findings. Most proteins were associated with a single cancer type, whereas 284 were associated with multiple cancers, indicating both shared and distinct biological pathways across cancer sites. Stratification by follow-up time revealed 523 protein associations that were restricted to specific time windows after blood collection (<2 years, <5 years, and >5 years), including 447 that were not apparent in primary analyses, highlighting proteins with potential roles as early detection or aetiological biomarkers. Proteins associated with cancer risk were enriched across diverse biological pathways and 315 corresponded to targets of approved cancer and non-cancer medications, while most proteins represent previously unrecognised links to cancer risk. As the most comprehensive prospective investigation of the plasma proteome and cancer risk to date, our study significantly expands our understanding of key biological pathways shaping cancer development and provides a resource for the scientific community to accelerate development of early detection strategies, biomarkers for risk stratification, and novel therapeutic interventions for cancer prevention.

Antigen-recognition histories accumulate across adult life, but existing human aging resources largely emphasize host-state measurements rather than antigen-resolved humoral memory. We present CHAIR, the Chinese Healthy Aging Immune Resource, which integrates self-antigen, human virome and toxin/virulence-factor PhIP-seq with routine clinical measurements and T/B/NK phenotyping in clinically screened adults aged 20-94 years. The resource framework includes 1,008 unique PhIP-seq samples, 995 clinical records, 977 T/B/NK records and 853 participants with all three layers. Across modeled spaces comprising 23,584 self-antigen targets, 210 virome targets and 1,258 virulence-factor source targets, most measured antibody-reactivity architecture was consistent with broad preservation within domains, whereas age- and sex-associated signals were concentrated in focal targets and modules. Together, CHAIR provides a traceable, antigen-resolved reference framework based on clinically screened adults for investigating focal humoral remodeling across healthy adult aging.

Autism is highly heritable, yet many cases remain genetically unsolved even after short-read sequencing (SRS). Long-read sequencing (LRS) increases sensitivity for pathogenic variant discovery, especially for complex structural variants (SVs) in repetitive regions missed by SRS. However, LRS largely relies on blood-derived DNA; whether saliva can support genome-wide LRS at population scale remains to be seen. We report the first systematic evaluation of saliva-derived DNA for genome-wide LRS in autism, benchmarking 105 new saliva samples against 68 new blood samples and integrating 189 previously published LRS blood samples to expand the autism cohort and improve ultra-rare pathogenic variant discovery. A matched blood-saliva quad provided direct comparisons of variant detection between sample sources. Saliva samples contained substantially more nonhuman DNA than blood (mean 10% versus 0.005% of reads), reducing mean effective genome coverage from an estimated 28X to 24X after removal of nonhuman reads, compared with 34X in blood samples. Despite this reduction, saliva and blood samples yielded comparable numbers of validated (22,591 versus 23,064) and rare (209 versus 213) SVs per sample, which was also observed in the quad when matching coverage. Saliva LRS identified four new pathogenic and likely pathogenic variants missed by SRS. We report variants affecting MECP2, DNMT3A, CHD8, LMBRD2, POGZ, and PHF3; four additional candidates with potential relevance to autism; and an increase of diagnostic yield of 9-18%. Saliva-derived DNA support blood-comparable discovery of pathogenic variants missed by SRS, including SVs, with increased sequencing coverage compensating for nonhuman DNA contamination and reduced assembly contiguity, thus showing saliva DNA collection as a feasible strategy for large-scale LRS studies of patients and population-scale sequencing of cohorts.

Zoonotic influenza A viruses pose constitute a significant persistent threat to global public health, yet quantitative approaches for assessing their cross-species transmission potential remain limited. Here, we present InfluProto, a deep learning model that introduces prototype learning into host prediction for influenza A virus, framing conceptualizing host adaptation as a continuous process rather than a discrete outcome. InfluProto embeds viral sequences into a continuous representation space anchored by human, swine, and avian prototypes, using cosine distances as a quantitative measure of host adaptation. InfluProto accurately discriminated among the three host classes, achieving recall and precision exceeding 0.97, and substantially outperformed conventional classification models models--particularly on challenging zoonotic strains, including the 2009 pandemic H1N1 (pdm09) lineage. Beyond host classification, the learned representations captured biologically meaningful host-adaptive signatures,; notably, with pdm09 viruses occupying occupied an intermediate position between human and swine clusters, consistent with their evolutionary origin. Analysis of simulated reassortment further revealed that the introduction of human-origin segments into swine influenza viruses reduced their distance to the human prototype, indicating that InfluProto can detect reassortment-associated shifts in host adaptation. To facilitate broad public useaccess, we have developed an interactive web service for InfluProto. Collectively, our work establishes a continuous framework for characterizing host adaptation and evaluating zoonotic risk in influenza A viruses, with potential implications for pandemic surveillance and preparedness.

Individuals with autism spectrum disorder (ASD) and co-occurring mood or psychotic disorders face substantially elevated suicide-attempt risk, but prior research is largely limited to population-level incidence and cross-sectional risk-factor studies, leaving unclear whether individual-level, prospective prediction is possible in this population. We conducted a retrospective prediction model study using Epic Cosmos electronic health record (EHR) data from 322,879 individuals with ASD and co-occurring mood or psychotic disorders to predict suicide attempts at the next clinical encounter using only information available at or before the preceding encounter. Five machine-learning algorithms were compared, with performance evaluated across discrimination, calibration, and clinical utility. All algorithms achieved strong discrimination using the full predictor set (AUROC 0.87-0.91). A parsimonious feature set (22 predictors) matched or exceeded full-model performance, with extreme gradient boosting (XGBoost) achieving the highest discrimination and standardized net benefit post-ablation (AUROC = 0.90; AUPRC = 0.34), though net benefit was comparable for elastic net and logistic regression. A prior suicide attempt was the strongest individual predictor. Decision-curve analysis indicated that model-guided decisions provided greater net benefit than treating all or no individuals across a broad range of clinically plausible thresholds. These findings demonstrate that near-term suicide-attempt risk can be predicted with high decision utility in this high-risk population using data already captured in routine care, suggesting a feasible basis for future EHR-based clinical decision-support tools.

ICD coding is the backbone of medical-record structuring, insurance payment, and clinical research. Today's best automatic coders are deep models, but they assume tens of thousands of labeled records, GPU servers, and machine-learning engineers. Many hospitals, especially small ones, have none of these. We study Chinese electronic medical records (EMRs) under a hard budget: 2,200 labeled records, CPU only, and fully interpretable decisions. Our method is lightweight and follows the way human coders work. First, codebook knowledge injection marks standard disease and operation names in the text as useful clues. Second, multi-view field weighting gives more weight to the clinical sections that matter most. Third, character n-gram TF-IDF features feed linear classifiers that score each candidate code. Fourth, a small set of clinical rules acts as a final safety check. Our main methodological finding is about this last step: naively applied rules do more harm than good (overall F1 drops from 0.8070 to 0.7995, and main-operation accuracy from 0.9727 to 0.9432). We therefore audit each rule on held-out data by counting how often it fixes a wrong prediction or breaks a correct one, and we remove rules that do more harm than good. This turns rule post-processing into a safety check that never hurts. On the CCL2026-Eval Task 6 dataset, our method reaches macro-F1 of 0.899 (main diagnosis) and 0.966 (main operation) in 5-fold cross-validation. A frequency-stratified analysis shows a clear long-tail limit: codes seen fewer than three times score zero F1. In a head-to-head test, a fine-tuned Chinese pretrained encoder scores lower than our method on all four tasks at this training size, while needing a GPU. Ablation tests show that the gains come mainly from field-weighted text. The framework trains in about 40 minutes on an ordinary CPU, and every prediction can be broken down into interpretable feature contributions.