Martin Gerbert Frasch

Early life adversity (ELA) is proposed here to bias precision weighting in specific information channels, up-weighting potential threat and predictable input while down-weighting evidence that disconfirms expectations of danger, rather than altering a single global gain. We ground this predictive-processing (PP) claim in developmental neuroscience: calibration occurs within critical and sensitive periods whose molecular brakes, together with self-confirming patterns of evidence sampling, can stabilize precision settings beyond the conditions that produced them, and susceptibility to calibration varies between individuals. We distinguish the parsimonious core claim (altered precision) from a stronger, optional claim (entrenched maladaptive priors) and specify behavioral, computational, and neural signatures that dissociate them. A mapping table links dimensions of early experience (threat, deprivation, unpredictability, rigid invariance, prenatal maternal distress, and supportive caregiving) to predicted precision channels, developmental windows, measurable indices, and disconfirming observations. Mechanistically, we propose the anterior cingulate cortex as a hub for allocating gain across channels within a distributed circuit that includes the amygdala, insula, orbitofrontal and ventromedial prefrontal cortex, striatum, thalamus, and the noradrenergic, cholinergic, and dopaminergic systems, and we state what each node contributes. Apparently contradictory findings, threat-related hyper-reactivity alongside reduced reward-prediction-error encoding, are read as one biased allocation of gain. The framework is held to a symmetry constraint (supportive input should confer faster precision re-tuning), states its boundary conditions, and yields falsifiable predictions and testable implications for exposure-based, pharmacological, and caregiving interventions.

Fetal-monitoring biomarkers for neonatal hypoxic-ischemic brain injury face a structural gap: the mechanistic ground truth that would label a training set — perfusion pressure, the moment of decompensation, the injury time course — cannot be measured at scale or ethically in human pregnancy or labor, and generative synthetic data carry no mechanistic labels. We address this with a mechanistic computational model of the fetal cardiovascular, autonomic, and metabolic response to controlled hypoxic stress, and use it to test how beat detection and acquisition fidelity alter the interpretation of fetal-heart-rate-variability (HRV) biomarkers. The model integrates these systems forward in time across antepartum development (gestational-age growth scaling) and intrapartum stress (umbilical-cord occlusions), emitting synthetic monitoring signals (fetal heart rate, RR intervals) co-registered with model-computed latent labels (pH, base deficit, lactate, perfusion pressure, decompensation and injury states). In a fetal-sheep-derived autonomic-loop configuration we report three results. First, a phase-accumulator beat detector shows that the apparently physiologic baseline HRV of an earlier build was largely a detector artifact, and a noise-off control shows beat-to-beat variability requires an explicit stochastic driver rather than self-sustained autonomic oscillation. Second, a sampling-fidelity sweep yields a fidelity-matched selection rule: a deceleration-area biomarker is preserved at CTG-grade 4 Hz whereas RMSSD is corrupted there (inflated about 8-fold by timing quantization) and recovers only at fetal-ECG rates. Third, autonomic modulation alone does not reproduce the published RMSSD rise-then-collapse — a negative result that motivates, but does not prove, an intrinsic sinoatrial-pacemaker contribution as a testable hypothesis. This is an in-silico, hypothesis-generating study: the model is not validated for individual fetal prediction, clinical risk estimation, or clinical decision-making. The model is implemented as Fetal Twin (engine fetaltwin ), a source-available research instrument released under a noncommercial license, together with all figure configurations, so that these controlled experiments are reproducible. Key Points: Progress on fetal-monitoring biomarkers for neonatal brain-injury risk is constrained by a structural gap: the mechanistic ground truth that would label a training set — perfusion pressure, the moment of cardiovascular decompensation, the time course of injury — cannot be measured at scale or ethically during human pregnancy or labor. We present Fetal Twin (source-available engine fetaltwin ), a publicly available, noncommercially-licensed mechanistic testbed for fetal physiological development. It integrates the fetal cardiovascular, metabolic, and autonomic systems forward in time across antepartum development (gestational-age growth scaling) and intrapartum stress (umbilical-cord occlusions), and emits synthetic monitoring signals (fetal heart rate, RR intervals) co-registered with model-computed latent labels (pH, lactate, perfusion pressure, decompensation and injury states). The name’s digital-twin connotation is deliberate but bounded: Fetal Twin is a mechanistic, population-level twin of fetal physiology used as a research instrument — not a validated, patient-specific clinical digital twin or predictor. Its purpose is to interrogate what candidate biomarkers can and cannot mean, via in-silico controls impossible in vivo — swapping the beat detector, turning a noise source off, ablating a reflex, or quantizing the signal to a monitor’s sampling grid. Demonstrations in a fetal-sheep-derived autonomic-loop configuration show that beat-to-beat HRV amplitude can be a numerical artifact of the beat detector, and that in this model class beat-to-beat variability requires an explicit stochastic driver — it does not arise as a self-sustained oscillation of the deterministic autonomic loop. A further demonstration establishes a fidelity-matched biomarker-selection rule — a deceleration-area biomarker survives CTG-grade 4 Hz sampling whereas RMSSD is corrupted at that rate and needs fetal-ECG timing — and a negative result shows the published RMSSD “rise-then-collapse” is not reproducible from autonomic modulation in this model class, motivating (but not proving) an intrinsic-pacemaker hypothesis.

Large language models in autonomous agents process trusted instructions and untrusted data in one context window, exposing them to direct and indirect prompt injection. In healthcare the stakes are concrete: a 2025 JAMA Network Open study found commercial medical LLMs followed injected instructions in 94.4% of simulated patient encounters, including life-threatening recommendations [21]. Yet the decisive problem is different: most real clinical threats—protected health information (PHI) exfiltration, cross-patient access, bulk export, out-of-scope advice—are fluent, legitimate-looking requests that carry no attack signal, so even a state-of-the-art injection detector passes them. We present QFIRE , an inline, provider-agnostic prompt firewall built as a single self-contained Rust toolchain (proxy, CLI, harness). QFIRE combines (i) positive-security scope constraints that confine a call to a declared natural-language purpose and block out-of-scope drift, (ii) an asynchronous detector graph that runs N rules concurrently, cheapest first, and (iii) a de-obfuscation pass (Base64/hex/ROT13, homoglyphs, leetspeak, zero-width). It ships 106 versioned rules and a HIPAA Safe-Harbor (18-identifier) PHI panel, and runs a local DeBERTa-v3 classifier via embedded ONNX Runtime. On 1,968 public prompts QFIRE attains F 1 0.86, statistically tied with Meta’s PromptGuard-2 (0.86) and above protectai DeBERTa-v3 (0.83). Our central result is on QFIRE-HealthBench , a new 2,000-prompt healthcare benchmark we release with real garak and Microsoft PyRIT [25] payloads: there PromptGuard-2 recovers only 0.40 recall while QFIRE’s combined scope+PHI chain reaches 0.83 ( F 1 0.87) at a calibrated 0.08 false-positive rate—generic injection detection is necessary but not sufficient for healthcare. A bare LLM judge matches QFIRE on static accuracy ( F 1 0.90) but falls to 34–59% recall under adaptive attack (§5.5), where QFIRE adds auditable determinism and bounded latency. End-to-end, placing QFIRE in front of a tool-using agent over a mock-EHR sandbox cuts its harmful-action rate from 0.38 to 0.00 at a 0.13 benign-utility cost. All code, rules, and corpora snapshots are released and regenerate from a single make paper target with no paid API keys.

HRV complexity metrics track the systemic cytokine response across mammalian inflammation models, but the dynamical mechanism translating molecular inflammation into HRV structure has remained unresolved, and no existing tool jointly extracts coordination-dynamics signatures from simultaneous vagus nerve electroneurogram (VENG) and ECG. We describe a modular open-source Python pipeline that (i) detects spectrally distinct latent states in 20 kHz whole-trunk VENG using a symmetric-Kullback–Leibler scan-statistic changepoint detector ported from earthquake seismology and verified against its R progenitor; (ii) places the inter-changepoint-interval series in the Haken–Kelso–Bunz coupled-oscillator framework, computes the Kuramoto order parameter R , fits the bistable potential V ( φ ) = − a cos φ − b cos 2 φ , and maps each timepoint into the Arnold-tongue (Ω, K ) plane; contrasts compound-action-potential band power between successive latent states; and (iii) couples the VENG analysis to a CIMVA HRV pipeline driven by a 5-algorithm Pro-MAC R-peak ensemble. Applied to a previously published two-animal neonatal piglet endotoxaemia cohort (Castel et al., 2020, 2024), the pipeline recovers known features of the (iv) preparation and surfaces three case-series observations that we frame as hypotheses for prospective replication: (1) a 6–8 s VENG state-switching period coincident with the HRV LF/HF spectral boundary; (2) qualitatively divergent coordination-dynamics trajectories under LPS versus LPS + VNS (Kuramoto R collapses from 0.69 to 0.12 without VNS, undergoes biphasic recovery to 0.99 with VNS, the two trajectories crossing at 60–75 min); (3) a regime-dependent empirical coupling between R and the HRV embedding scaling exponent eScalE ( R · eScalE ≈ 0.33 during moderate challenge, doubling to ≈0.80 during VNS overshoot). Pre-specified null comparisons (Poisson and bootstrap surrogates) demonstrate that R and the critical-slowing-down indicators carry information beyond their most obvious null alternatives. With N = 2 animals no inferential statistics are possible, so these findings are reported as hypothesis-generating; the manuscript closes with a pre-registered prospective protocol designed specifically to falsify them. Graphical Key Points: We present an open-source Python pipeline that extracts coordination-dynamics invariants (Kuramoto phase coherence, Haken–Kelso–Bunz potential parameters, Arnold-tongue position, critical-slowing-down indicators) from simultaneously recorded vagus nerve electroneurogram (VENG) and ECG, starting from a spectral-KL scan-statistic changepoint detector ported from earthquake seismology. Applied to a previously published two-animal cohort (Castel et al., 2024), the pipeline reveals that VENG alternates between two spectrally distinct latent states at a baseline period of 6–8 s that coincides with the HRV low-frequency/high-frequency spectral boundary (∼0.15 Hz). We propose the VENG latent-state oscillator as a candidate mechanistic generator of vagal LF HRV; this is a hypothesis motivated by the timescale match and is not established by the present data. In this cohort, LPS endotoxaemia and vagus nerve stimulation drove qualitatively divergent trajectories in the coordination-dynamics phase space (Kuramoto R : 0.69 → 0.12 without VNS; biphasic to R → 0.99 with VNS at the IL-6/IL-8 cytokine peak). Because the design provides one animal per arm, these are case-series observations and not a between-arm statistical comparison. Inter-changepoint-interval variance rose >100-fold roughly 15 min before overt state-switching collapse in the unmitigated animal. Bootstrap nulls drawn from the same animal’s pre-collapse ICI pool reject the same-distribution alternative (§2.6), motivating a critical-slowing-down early-warning hypothesis that requires prospective replication before any clinical use is warranted. The pipeline surfaces a regime-dependent empirical coupling between VENG phase coherence ( R ) and HRV embedding scaling exponent (eScalE) in the VNS-treated animal: during moderate challenge R · eScalE ≈ 0.33 (CV ≈ 0.20), during VNS-enhanced overshoot (45, 90, 105 min) the product doubles to ≈0.80. The product is a post-hoc empirical construct, not a derived conservation law; we propose it as a falsification target for prospective cohorts (§6).

Prenatal maternal stress alters maternal–fetal heart rate coupling, as demonstrated by the Fetal Stress Index derived from bivariate phase-rectified signal averaging. Here, we extend this framework using information-theoretical measures to elucidate underlying mechanisms. In 118 third-trimester pregnancies (59 stressed, 59 control), we computed entropy rate (ER), sample entropy (SE) and transfer entropy (TE) under multiple conditioning paradigms, employing mixed linear models for repeated measures. We identified dual coupling mechanisms at the short-term (0.5–2.5 s), but not long-term (2.5–5 s) time scales: (1) stress-invariant state-dependent synchronization, with maternal decelerations exerting approximately 60% coupling strength on fetal heart rate complexity—a fundamental coordination conserved across demographics; and (2) stress-sensitive temporal information transfer (TE), showing exploratory associations with maternal cortisol that require replication. A robust sex-by-stress interaction emerged in TE from mixed models, with exploratory female-specific coupling patterns absent in males. Universal acceleration predominance was observed in both maternal and fetal heart rates, stronger in fetuses and independent of sex or stress. We provide insight into the dependence of these findings on the sampling rate of the underlying data, identifying 4 Hz, commonly used for ultrasound-derived fetal heart rate recordings, as a sufficient sampling rate to capture the information flow. Information-theoretical analysis reveals that maternal–fetal coupling operates through complementary pathways with differential stress sensitivity, extending the Fetal Stress Index by elucidating causal foundations. Future studies should explore additional information-theoretical conditional approaches to resolve stress-specific and time-scale-specific differences in information flow.

As health systems begin deploying autonomous AI agents that make independent clinical decisions and take direct actions within care workflows, ensuring patient safety and care quality requires governance standards that go beyond existing medical device frameworks designed for human-in-the-loop prediction tools. This paper introduces the Healthcare AI Agents Regulatory Framework (HAARF), a comprehensive verification standard for autonomous AI systems in clinical environments, developed collaboratively with 40+ international experts spanning regulatory authorities, clinical organizations, and AI security specialists. HAARF synthesizes requirements from nine major regulatory frameworks (FDA, EU AI Act, Health Canada, UK MHRA, NIST AI RMF, WHO GI-AI4H, ISO/IEC 42001, OWASP AISVS, IMDRF GMLP) into eight core verification categories comprising 279 specific requirements across three risk-based implementation levels. The framework addresses critical gaps in health system readiness for autonomous AI, including: (1) progressive autonomy governance with clinical accountability, (2) tool-use security for agents that independently access EHRs, medical devices, and clinical systems, (3) continuous equity monitoring and bias mitigation across diverse patient populations, and (4) clinical decision traceability preserving human oversight authority. We validate HAARF's enforcement capabilities through a scenario-based red-team evaluation comprising six adversarial scenarios executed under baseline (no middleware) and HAARF-guardrailed conditions (N = 50 trials each, Gemini 2.5 Flash primary with Claude Sonnet 4.6 cross-model validation). In baseline conditions, the agent model executes unauthorized tools in 56-60% of adversarial trials. Under the HAARF condition, deterministic middleware enforcement reduces the unauthorized-tool success rate to 0%, with 0% contraindication misses and 0% policy-injection success (95% Wilson CI [0.00, 0.07]). Cross-model validation confirms identical security metrics, supporting HAARF's model-agnostic design. Mapping analysis demonstrates 48-88% coverage of major regulatory frameworks, with per-category FDA alignment ranging from 73% (C5, Agent Registration) to 91% (C3, Cybersecurity; C7, Bias and Equity). Initial validation with healthcare organizations shows a 40-60% reduction in multi-jurisdictional compliance burden and improved clinical safety governance outcomes. HAARF provides health systems with a practical, risk-stratified pathway for safe AI agent deployment - shifting from reactive compliance to proactive quality governance while maintaining rigorous patient safety standards and human-centered care principles.

BBackground: Postpartum depression (PPD) affects 10-20% of women following childbirth yet remains systematically underdiagnosed, with screening and treatment rates varying by race, ethnicity, and geography. Machine learning models for PPD risk stratification risk perpetuating these disparities if not evaluated for algorithmic bias across intersecting demographic dimensions. We developed a graded-severity PPD prediction model on the largest US population-based dataset, characterised its demographic biases, and evaluated whether geospatial foundation model embeddings could reduce intersectional disparities while improving overall performance.Methods: Using data from 209,872 respondents in the CDC Pregnancy Risk Assessment Monitoring System (PRAMS), we developed gradient boosting classifiers for five-class PPD severity prediction (Always/Often/Sometimes/Rarely/Never). We performed stratified bias analysis across race (seven groups), ethnicity, urbanicity, and their intersections. We then enhanced predictions by integrating 330-dimensional geospatial embeddings from Google's Population Dynamics Foundation Model (PDFM), aggregated at the state-by-Rural-Urban Commuting Area (RUCA) level across four geographic stratification strategies.Findings: The baseline model achieved 82.5% sensitivity and AUC 0.673 for very high risk (Always) detection but exhibited substantial intersectional disparities: a 9.0 percentage-point urban-rural gap, a 2.5 percentage-point Hispanic disparity, and a 31.2 percentage-point range across race-by-urbanicity intersections. The PDFM-enhanced model with 4-category RUCA stratification improved AUC to 0.719 (+6.8%) and sensitivity to 88.3% (+5.8 percentage points), while reducing the urban-rural gap by 89% (to 1.0 percentage points) and eliminating the Hispanic disparity by 96% (to 0.1 percentage points). Black women's sensitivity improved from 86.2% to 92.0% and American Indian/Alaska Native women's from 91.7% to 95.3%.Interpretation: Geospatial foundation model embeddings can simultaneously improve predictive performance and reduce intersectional demographic disparities in PPD screening. This contextual enrichment approach offers a scalable template for building equitable clinical prediction tools.

Objective: To evaluate modifiable antepartum and intrapartum factors associated with nulliparous, term, singleton, vertex (NTSV) cesarean delivery and to model risk stratified induction timing strategies that minimize cesarean risk across maternal risk profiles. Study Design: This retrospective cohort study included all NTSV deliveries at a tertiary care center from January 2015 through August 2025 (overall cohort n=10,525; limited risk cohort n=5,663). Machine learning identified key predictors of cesarean delivery, with maternal age and pre pregnancy body mass index (BMI) used to define low, moderate, and high risk strata. Logistic regression estimated the association between induction and cesarean delivery, and a Monte Carlo simulation compared elective induction at 39, 40, or 41 weeks versus expectant management to 42 weeks within each stratum. Results: Cesarean delivery occurred in 20.1% of the overall cohort and 19.0% of the limited risk cohort, with a U shaped relationship between gestational age and cesarean risk and lowest rates at 38-39 weeks. Induction was associated with higher cesarean rates than spontaneous labor in both cohorts (overall: 24.1% vs. 17.1%; limited risk: 22.9% vs. 15.7%) after adjustment for age, BMI, and gestational age. No single induction policy minimized cesarean risk across all strata. For high risk patients (age >=35 years and BMI >=35), induction at 39 weeks yielded the lowest modeled cesarean rate, whereas later delivery (40 to 41 weeks or expectant management to 41 weeks) was favored for low and moderate risk patients. A universal 39 week induction policy for low and moderate risk strata modestly increased modeled cesarean rates, adding an estimated 46 cesarean deliveries. Conclusion: Gestational age at delivery and induction strategy are key modifiable determinants of NTSV cesarean delivery, but optimal timing varies by maternal age and BMI risk profile, supporting risk stratified rather than universal 39 week induction policies.