Stress Simulators as Negative Controls: Correcting Selection Bias in Retrospective Metabolic Cohorts

When researchers study how chronic metabolic stress — years of elevated blood glucose, fatty acids, and insulin resistance — damages organs, they face a cruel trade-off. Wet-lab simulations compress decades of disease into days by bathing cells in extreme nutrient conditions, which isolates the biochemistry but throws away the real clinical world. Retrospective clinical databases capture the real world, but only through a badly distorted lens: patients are tested when doctors are worried, treated when doctors are more worried, and included in cohorts only if they survived, returned, and had complete records. The result is that the two literatures rarely constrain each other. This proposal argues they should: a mechanistic simulator of glucose-driven tissue injury, calibrated to in vitro stress experiments, can act as a structural negative control that detects and partially corrects selection bias in retrospective metabolic cohorts — something no purely statistical method can do, because statistics alone cannot tell you what the biology should look like.

The concrete motivating case is documented in the file. In a retrospective cohort of 26,197 Type 2 diabetes patients under primary care in Hong Kong, patients prescribed ACE inhibitor or ARB drugs showed faster kidney disease progression, which the authors correctly identified as confounding by indication: doctors gave renoprotective drugs to patients who were already deteriorating An individualized risk prediction model for new-onset, progression and regression of chronic kidney disease in a retrospective cohort of patients with type 2 diabetes under primary care in Hong Kong. The same cohort required complete eGFR, UACR, and HbA1c records to enter the analysis at all, so the sickest, least-monitored patients vanished. Neither problem was corrected; both were acknowledged and set aside. Meanwhile, in vitro work has established precisely the dose-response machinery that the retrospective analysis lacks: hyperglycemia drives reactive-oxygen-species accumulation, antioxidant depletion (GPX, GPX4), p53/p21 cell-cycle arrest, and eventually irreversible structural damage, along the AMPK–SIRT1–NRF2 and NRF2/HO-1 axes Aloperine attenuates high glucose-induced oxidative injury in Schwann cells via activation of NRF2/HO-1 pathway, FGF4 alleviates the lung cell damage caused by high glucose via AMPK‑PGC‑1 signaling axis in vitro and in vivo, FGF21-engineered ADSCs promote diabetic wound healing by mitigating ferroptosis and oxidative stress via the SIRT1/NRF2/GPX4 signaling pathway. My claim is that this machinery, expressed as a compact dynamical model, is exactly the prior needed to make retrospective metabolic data causally interpretable.

Core Claim

Retrospective selection bias in metabolic cohorts can be detected and reduced by fitting a joint model in which a mechanistic stress-injury simulator generates a latent tissue-damage trajectory, an explicit selection model generates the clinical actions (test ordering, drug initiation, cohort inclusion) that depend on that latent, and the observed irregular labs are noisy, selectively sampled readouts. The key testable implication is asymmetry: selection bias distorts the observed exposure–injury relationship away from the mechanistically predicted one in patterns that correlate with testing and treatment behavior, and correcting for the selection mechanism should restore a dose-response consistent with in vitro biology and with known protective interventions.

The closest existing work, DTD-Sim, fuses the FDA-accepted UVA/Padova glucose-insulin ODEs with a deep state-space model so that physiological parameters (insulin sensitivity, carb absorption) become time-varying latent states rather than constants; it beat an LSTM by 36% in mean absolute error at a 6-hour forecast horizon on two participants' wearable data, and produced physically sensible counterfactual glucose spikes where unconstrained networks produced impossible plateaus Learning Insulin-Glucose Dynamics in the Wild. DTD-Sim is the methodological ancestor of this proposal, but it assumes logged insulin and meals are accurate, ignores adherence and reporting bias entirely, models nothing below the organ level, and has no notion of clinical selection — with only two participants, every observation was effectively self-selected by the user. The other closest work, the Fair Longitudinal Medical Deconfounder, infers latent confounders from longitudinal records with a VAE–LSTM and enforces counterfactual fairness, showing that statistically inferred latents can absorb unobserved confounding on MIMIC-III and NSQIP A Counterfactual Fair Model for Longitudinal Electronic Health Records via Deconfounder. But its latents are validated only by reconstruction loss — the paper itself does not ground them in any biological mechanism — and they have no dynamics that constrain what a decades-long damage trajectory can look like. My proposal occupies the empty corner: mechanistic latents with explicit dynamics, coupled to an explicit model of who gets measured and treated.

Research Program

The model has three parts. First, a damage accumulator, the mechanistic core, distilled from the in vitro literature. Write the latent tissue damage $D_t$ at time step $t$ as

$$D_{t+1} = D_t + \alpha\,\sigma(g_t - g^*) - \rho(D_t)\,D_t,$$

where $g_t$ is ambient glycemic exposure, $\sigma$ is a sigmoid stress response, $g^*$ is a stress threshold, $\alpha$ scales injury inflow, and $\rho(D_t)$ is a repair rate that decays toward zero once $D_t$ crosses an irreversibility threshold $D^*$. This functional form is not invented for convenience: it is the systems-level reading of the established biology, in which hyperglycemic ROS generation exceeds antioxidant capacity (GPX/HO-1 depletion) Aloperine attenuates high glucose-induced oxidative injury in Schwann cells via activation of NRF2/HO-1 pathway and damage becomes self-sustaining through senescence-associated secretory signaling and ferroptotic lipid peroxidation FGF4 alleviates the lung cell damage caused by high glucose via AMPK‑PGC‑1 signaling axis in vitro and in vivo, FGF21-engineered ADSCs promote diabetic wound healing by mitigating ferroptosis and oxidative stress via the SIRT1/NRF2/GPX4 signaling pathway. It also directly mirrors the strongest empirical finding in the retrospective literature: in the Hong Kong cohort, the probability of regression from a chronic kidney disease stage drops toward zero after two years of residence in that stage, across all ages and sexes — a clean observational signature of $\rho \to 0$ An individualized risk prediction model.... The parameters $\alpha$, $g^*$, and $D^*$ are patient-specific latent states with their own slow drift, exactly as DTD-Sim lets insulin sensitivity drift Learning Insulin-Glucose Dynamics in the Wild. Crucially, the in vitro dose-response curves calibrate the shape of the stress term (monotone, thresholded), not its absolute scale: a 150 mM Schwann-cell shock over 48 hours and a decade of 10 mM ambulatory glucose differ in magnitude, and the model's job is to bridge that temporal compression with learned scale factors rather than pretend the dish equals the patient — the "physiological complexity gap" that the reductionist literature itself flags.

Second, a selection model. Clinical actions are generated as logistic functions of the latent $D_t$ and recent observed signals: the probability that a glucose or creatinine measurement is ordered, that a renoprotective drug is initiated, that a patient returns for follow-up, and that a patient passes cohort inclusion filters. This treats test ordering as informative, in the spirit of work showing that the timing and mere presence of a measurement carries clinical signal about acuity Improving Medical Predictions by Irregular Multimodal Electronic Health Records Modeling, but inverts it: rather than exploiting informative timing for prediction accuracy, we model the mechanism that produced the timing so we can divide it out of causal estimates.

Third, joint estimation. Inference alternates between fitting the damage trajectory (given observed labs and actions) and fitting the selection parameters (given the inferred latents), in the manner of the two-stage deconfounder A Counterfactual Fair Model for Longitudinal Electronic Health Records via Deconfounder. The correction is then a straightforward inverse-probability step, but with weights that depend on the mechanistic latent, not merely on observed covariates — which is precisely what standard propensity methods cannot do, since the confounder driving treatment assignment is the latent severity itself.

Proposed Experiments

The central difficulty in evaluating any debiasing method is that ground-truth causal effects in observational data are unknown. I therefore build the evaluation in three tiers of decreasing control.

The first tier is semi-synthetic and uses MIMIC-III through the MIMIC-Extract pipeline, which standardizes units, clips physiologically impossible values, and maps thousands of raw items into 104 clinical aggregates including blood glucose, creatinine, and lactate MIMIC-Extract: A Data Extraction, Preprocessing, and Representation Pipeline for MIMIC-III. I will use the mechanistic simulator to generate synthetic latent damage trajectories and lab values for real patient timelines, then impose the documented selection filters — the pipeline's own 12-hour-to-10-day stay restriction, informative test-ordering learned from the data, and complete-record inclusion — and hide a known ground-truth exposure–injury coefficient and a known protective drug effect inside the simulation. Each estimator (naive, propensity-based, deconfounder, ours) must recover them. Metrics are absolute bias of the recovered dose-response coefficient, 95% confidence-interval coverage, and calibration of inferred severity, plus predictive area under the ROC curve on a held-out final window, following DTD-Sim's protocol of training on early data and testing on the last stretch. The honest framing for the real-data reference point: MIMIC-Extract reports AUROCs between 0.85 and 0.90 for mortality prediction and macro AUROCs of 0.90 to 0.94 for intervention onset/offset, with area under the precision-recall curve often below 0.50 due to class imbalance MIMIC-Extract — useful context for what predictive difficulty looks like on this database, though prediction is not our endpoint.

The second tier is a real-data falsification test on the chronic metabolic endpoint. The Hong Kong study is the anchor: its multi-scale multi-state Poisson model tracks new-onset, progression, and regression of diabetic kidney disease with progression AUC of 0.72–0.84 and regression AUC of 0.50–0.71, and it flags uncorrected confounding by indication for ACEI/ARB prescription An individualized risk prediction model.... Where cohort access is possible, I replicate the analysis; where not, I construct an analogous T2DM cohort from accessible longitudinal records with the same eGFR/UACR staging. The decisive check is directional: after selection-aware correction, the apparent harm of ACEI/ARB initiation should attenuate or reverse (renoprotection is one of the best-established treatment effects in nephrology), while a naive complete-case re-analysis should reproduce the spurious harm. If correction does not move the estimate, the proposal's key ingredient has failed on its home turf.

The third tier is a translation check against in vitro biology. The corrected in-vivo dose-response must be monotone in glycemic burden with a threshold, and known protective compounds must shift it in the right direction with the right pathway signature — aloperine's NRF2/HO-1 activation reversing high-glucose injury in Schwann cells at 1–50 μM against a 150 mM glucose challenge Aloperine attenuates high glucose-induced oxidative injury in Schwann cells..., FGF4's AMPK–PGC-1α rescue of lung epithelium under 30 mmol/L glucose FGF4 alleviates the lung cell damage..., and FGF21-expressing stromal cells preventing GPX4-loss ferroptosis and achieving nearly 100% wound closure by day 14 in diabetic mice FGF21-engineered ADSCs.... These are qualitative anchors, not numeric targets, and I will state them as such; the file gives no quantitative in-vitro/in-vivo bridging number, and I will not manufacture one. No reported result found on Lacuna for a calibrated quantitative mapping between in-vitro glucose dose-response and human cohort outcomes — that gap is itself part of what this work fills.

Baselines

The naive reference point is complete-case retrospective regression: fit a standard multivariable survival or multi-state model on whatever patients pass the inclusion filters, with no treatment of missingness or treatment assignment. Concretely this is the Hong Kong multi-scale multi-state approach itself, which is both the strongest published model on this endpoint and, uncorrected, the embodiment of the bias — it reports progression AUC of 0.72–0.84 and regression AUC of 0.50–0.71 An individualized risk prediction model.... It will be run on identical cohorts, features, and event definitions, with the same tuning budget, and it is the strongest competitor for prediction; beating it on causal recovery is the actual target, since we should not expect to beat it on AUC and will say so.

The strongest existing debiasing method is the Fair Longitudinal Medical Deconfounder: a VAE–LSTM that infers per-encounter latent confounders from the observed sequence, then predicts with those latents appended A Counterfactual Fair Model for Longitudinal Electronic Health Records via Deconfounder. It will be retrained on the same cohorts with a matched parameter count and the same cross-validation folds. It wins if its statistically inferred latents absorb the severity confounding as well as my mechanistic latent — this is the comparison that decides the paper. The file reports no numerical MIMIC-III benchmark value for FLMD beyond comparative descriptions, so head-to-head numbers must come from our own runs under matched conditions.

The key ablation is the simulator-without-selection variant: the same mechanistic damage accumulator and state-space estimation as my full model, but with test ordering and treatment treated as ignorable (measurements modeled as missing at random). This isolates the selection mechanism as the proposal's active ingredient; if it performs as well as the full model in the semi-synthetic experiment, the selection component is not earning its place and the idea reduces to DTD-Sim-style mechanistic forecasting, which the bias-recovery metrics should refute. I will also run propensity-score matching and inverse-probability weighting on observed covariates, and an irregular-time predictive model with time-aware embeddings Improving Medical Predictions by Irregular Multimodal Electronic Health Records Modeling as a strong predictive-but-noncausal reference, both on identical data and compute.

Why not just…?

Why not just use propensity scores or inverse-probability weighting? These are the standard tools, and they are cheap. But they adjust only for observed covariates, and the documented failures are exactly about the unobserved: SEER-based analyses of gastric signet ring cell carcinoma remain blind to performance status, subclinical reserve, and socioeconomic barriers after full observed-covariate adjustment, and the Hong Kong ACEI/ARB case shows observed labs failing to remove spurious treatment harm because the confounder is the doctor's read of deterioration that precedes the labs An individualized risk prediction model.... My semi-synthetic experiment makes this quantitative: propensity methods are included as baselines and will be scored on bias recovery where the true confounder is latent by construction. I expect them to underperform, and if they do not, that result will be reported.

Why not just use the deconfounder (FLMD) without any mechanistic simulator? This is the real threat, and I will not argue around it. FLMD already showed that VAE-inferred latents absorb unobserved confounding and improve fairness metrics on MIMIC-III and NSQIP A Counterfactual Fair Model for Longitudinal Electronic Health Records via Deconfounder. My honest answer is two-part. First, its latents carry no dynamics and no biological grounding, so they cannot enforce the monotone, thresholded, eventually-irreversible shape that damage trajectories must have — a constraint with real empirical teeth in the two-year irreversibility finding — and ungrounded latents cannot support extrapolation to the counterfactual exposure regimes that retrospective correction requires. Second, and more decisive: this is an empirical question, and FLMD is run head-to-head on the identical semi-synthetic benchmark with matched capacity. If grounded latents add nothing over free latents, the mechanistic half of the proposal is decoration, and I will conclude that.

Why not just collect better data — wearables or a prospective trial? Prospective randomization would indeed dissolve selection bias, but it cannot rerun thirty years of metabolic exposure, and the endpoint that matters (irreversible organ injury) is measured in decades. Wearables help at the systemic level, and DTD-Sim shows how much: a 36% MAE improvement over LSTMs at a 6-hour horizon — but on two participants who logged their own meals and insulin, with adherence and reporting bias explicitly out of scope Learning Insulin-Glucose Dynamics in the Wild. Self-selected logging is itself a selection mechanism, and wearables say nothing about clinical test-ordering or cohort inclusion in hospital data. The realistic division of labor is that wearables supply dense exposure measurement for a lucky few, while retrospective EHRs remain the only record of long-run outcomes for everyone — which is why correcting them, rather than abandoning them, is worth the effort.

Decisive Checks

The proposal stands or falls on three outcomes. In the semi-synthetic MIMIC-III experiment, the full model must recover the planted exposure–injury and treatment-effect coefficients with materially lower bias and correct coverage than propensity baselines, the ablation without the selection head, and a capacity-matched deconfounder. On the real chronic-kidney endpoint, the corrected ACEI/ARB estimate must move away from spurious harm while the naive replication reproduces it. And the corrected dose-response must be monotone, thresholded, and eventually irreversible — the shape the in vitro stress literature has already established in the dish Aloperine attenuates high glucose-induced oxidative injury in Schwann cells..., FGF4 alleviates the lung cell damage.... Any one of these failing names the component at fault; all three passing would establish, for the first time in this literature, a two-way street between the compression of the cell culture and the distortion of the medical record.