Inflammation History Predicts How Muscle Fails

When a person can no longer hold a contraction, the failure can come from several places: the brain may send less drive down to the spinal cord, the spinal reflex loop may change its gain, or the muscle fibers' own membranes may conduct electrical impulses more slowly as ion gradients run down. Surface electromyography — electrodes on the skin that pick up the electrical storm of contracting muscle — can tell these apart: the speed at which the muscle-fiber electrical wave travels (conduction velocity) indexes the peripheral, membrane-side failure, while evoked spinal and cortical responses index central drive. Separately, a large geriatric cohort showed that a cheap inflammation marker from a routine blood count — the ratio of neutrophils to lymphocytes — predicts who will become frail two years later, with an odds ratio of 2.92 for the most affected individuals in Association of immunity markers with the risk of incident frailty: the Rugao longitudinal aging study. But nobody knows whether immune and inflammation history predicts how a person's muscles fail, only whether they fail on a gross behavioral test.

This note proposes exactly that linkage, and proposes to build it where the immune data actually lives: in retrospective electronic health records. The design is a hybrid. We recruit adults from a health system with multi-year records, reconstruct each person's longitudinal inflammation and immune-exposure trajectory from historical blood counts and encounter data, and then bring them in for a single laboratory session in which surface electromyography resolves their personal fatigue mechanism. Retrospective cohort machinery — propensity weighting, target-trial-style covariate balance — does the causal hygiene. The payoff is a mechanistic bridge between two literatures that have never shared a cohort: exercise physiology, which shows training reshapes central and peripheral fatigue pathways in samples of eight to twenty-five men, and immunology, which shows those same pathways' upstream systemic milieu drifting over years.

Core Claim

The claim is that a person's multi-year inflammation trajectory — computable from routine complete blood counts already sitting in the electronic health record — predicts which fatigue mechanism dominates under a standard sustained contraction, and that this prediction is invisible to gross functional tests and to single-baseline inflammation snapshots. We further expect the relationship to be moderated by training history: the exercise literature gives reason to think endurance-style training buffers the peripheral (conduction-velocity) pathway specifically, and if chronic inflammation attacks that same pathway, trained and untrained people with identical inflammation histories should fail differently.

Three strands of existing work motivate this but stop short. The training studies of Vila-Chã and colleagues show that endurance and strength training tune different parts of the neuromuscular system: after three weeks, endurance training raised the spinal reflex ratio $H_{\max}/M_{\max}$ by 30.8 ± 21.7% while strength training raised the cortical-drive ratio $V/M_{\max}$ by 55.1 ± 28.3% in Changes in H reflex and V wave following short-term endurance and strength training, and after six weeks endurance training cut the rate of conduction-velocity decline during a sustained 30% maximal contraction while also reducing antagonist coactivation by about 37%, per Adjustments in Motor Unit Properties during Fatiguing Contractions after Training. These studies supply the mechanistic phenotype but use eight to twenty-five healthy sedentary men, no immune data, and windows of weeks. The Rugao study supplies the immune-prediction logic but measures only behavioral frailty (grip strength, walking speed, exhaustion questionnaires) and — importantly — found its blood-count ratios predicted incident but not prevalent frailty, so a single snapshot says little about current physiological state. The Long COVID cohort work, Immunological Density Shapes Recovery Trajectories in Long COVID and Real-World Effectiveness of Nirmatrelvir in Protecting Long COVID for Outpatient Adult Patients – A Large-Scale Observational Cohort Study from the RECOVER Initiative, supplies the retrospective-EHR methodology at scale, but measures fatigue only through symptom scores and billing codes; the first of these found that elapsed time alone does not resolve post-viral fatigue (correlation r = 0.0521) while cumulative vaccination does (r = -0.0434), which suggests the immune system's state actively shapes fatigue phenotypes rather than passively accompanying them. No reported result found on Lacuna for any study linking surface-electromyographic fatigue metrics to longitudinal immune trajectories within a retrospective cohort; that is the gap this proposal occupies.

Research Program

The cohort. We will recruit roughly 300 adults aged 40–75 from a single health system with at least five years of continuous records, sampling across quartiles of the inflammation trajectory (the slope and variance of the neutrophil-to-lymphocyte ratio and the systemic immune-inflammation index over all recorded blood counts, following the marker definitions of the Rugao study) and across self-reported and record-confirmed exercise-modality histories. A sub-cohort of about 60 participants, balanced across inflammation strata, will additionally give one blood draw for neutralizing-antibody titers and flow-cytometric memory B- and T-cell phenotyping, replicating the assay panel of Longitudinal Analysis of Humoral and Cellular Immune Response up to 6 Months after SARS-CoV-2 BA.5/BF.7/XBB Breakthrough Infection and BA.5/BF.7-XBB Reinfection at a single time point. This sub-cohort is a validation resource, not the discovery engine; its purpose is to check whether the cheap blood-count trajectories we rely on actually track the humoral-versus-cellular divergence documented there (antibodies waning steeply while T-cell responses stay stable and memory B cells expand) and in Immune Persistence against SARS-CoV-2 after Primary and Booster Immunization in Humans: A Large-Scale Prospective Cohort Study, where wild-type neutralizing titers fell from a geometric mean of 17.95 at three months to 8.00 at nine to twelve months.

The mechanistic fingerprint. In one 90-minute session, each participant performs the sustained 30%-of-maximum knee-extension contraction used by Adjustments in Motor Unit Properties during Fatiguing Contractions after Training, instrumented with the flexible adhesive multichannel electrode arrays validated for dynamic conduction-velocity estimation in Assessment of average muscle fiber conduction velocity from surface EMG signals during fatiguing dynamic contractions, where the multichannel method reached a root-mean-square error of about 0.2 m/s versus 0.8 m/s for two-channel arrays — which is why we use arrays, not pairs. From the recordings we extract three feature families: the rate of conduction-velocity decline (peripheral mechanism), motor-unit recruitment-threshold dynamics under decomposition as in Adjustments Differ Among Low-Threshold Motor Units During Intermittent, Isometric Contractions, which showed units active more than 90% of a task behave oppositely to units active under 70%; and evoked H-reflex and V-wave responses for the central split, per Changes in H reflex and V wave following short-term endurance and strength training. Each participant is assigned a continuous two-axis fingerprint: peripheral-share and central-share of their fatigue.

The analysis. The primary model asks how much of the variance in the peripheral-share axis is explained by inflammation-trajectory features, with training-modality history and its interaction entered as moderators, under inverse-probability-of-treatment-weighting balance of recorded confounders in the style of Real-World Effectiveness of Nirmatrelvir in Protecting Long COVID for Outpatient Adult Patients – A Large-Scale Observational Cohort Study from the RECOVER Initiative, which balanced nearly 500,000 patients across 29 record systems with this machinery. We will pre-register the analysis, run negative-control outcomes (associations with a variable the mechanism cannot plausibly touch, such as incident dermatologic diagnoses), and report how much unmeasured confounding would be needed to erase the effect.

Baselines

The naive reference point is the gross functional endpoint: maximal voluntary contraction and time-to-task-failure, exactly as Changes in H reflex and V wave following short-term endurance and strength training reports them (strength training raised maximal contraction by 14.4 ± 5.2% while endurance training raised time-to-task-failure by 22.7 ± 17.1%). This is the default measurement in clinical and geriatric settings, and it is the strongest competitor for the prediction-only question because it costs nothing. It will be collected in the same session on the same participants, and any model we build must beat a classifier built on these endpoints alone, fit on identical data with identical cross-validation folds and a matched hyperparameter search budget. Beating it means a pre-specified improvement of at least 0.08 in area under the receiver-operating curve for classifying the dominant fatigue mechanism, with a bootstrap confidence interval excluding zero — a bar that gross endpoints should find hard to clear, since the cited paper shows they can move in opposite directions under interventions that target different mechanisms.

The strongest existing method is the Rugao-style baseline-inflammation model: a logistic regression on a single blood-count snapshot (log-transformed neutrophil-to-lymphocyte ratio and systemic immune-inflammation index) predicting physical decline, which achieved an odds ratio of 2.92 (95% CI 1.20–7.18) for incident frailty in Association of immunity markers with the risk of incident frailty: the Rugao longitudinal aging study? — no; the correct identifier is Association of immunity markers with the risk of incident frailty: the Rugao longitudinal aging study, and the quoted odds ratios are 2.92 (95% CI 1.20–7.18) for log neutrophil-to-lymphocyte ratio and 2.34 (95% CI 1.16–4.78) for log systemic immune-inflammation index. We will run this exact specification on our cohort, predicting our mechanistic fingerprint instead of the Fried phenotype, with the same covariate set and the same tuning budget as our trajectory model. Beating it means the multi-year trajectory features (slopes, variance, immune-exposure density) add discrimination that a baseline snapshot cannot, by the same 0.08 area-under-the-curve margin, plus correct mechanism attribution that an odds ratio structurally cannot provide.

The key ablation isolates the mechanistic ingredient of the electromyography itself: two models with identical inputs (inflammation trajectory plus behavioral endpoints), one using only conventional amplitude-based electromyographic features (root-mean-square levels), the other using the decomposition and conduction-velocity features. If amplitude features do as well, the proposal's central claim about mechanism resolution collapses into "any muscle signal helps," and we would report that. As a secondary reference for what modern machine learning on surface electromyography achieves on related tasks, the gesture-fatigue long-short-term-memory model of sEMG-Based Neural Network Prediction Model Selection of Gesture Fatigue and Dataset Optimization reported a mean squared error of 0.0148 on wavelet coefficients, and the vemg2pose benchmark model of emg2pose: A Large and Diverse Benchmark for Surface Electromyographic Hand Pose Estimation reported 12.2° mean joint-angular error on held-out users; these anchor what is achievable on large surface-electromyography datasets and will be used to sanity-check our recording and modeling pipeline, not as direct competitors on a different task.

Why not just…?

Why not just use grip strength and walking speed? Because the motivating dissociation is already documented: strength training improved maximal contraction by 14.4% with no change in time-to-task-failure, while endurance training improved time-to-task-failure with no change in maximal contraction, and the two modalities moved the spinal reflex ratio and the cortical-drive ratio in opposite patterns (Changes in H reflex and V wave following short-term endurance and strength training). A dynamometer treats fatigue as one number and cannot say whether exhaustion came from falling central drive or slowing muscle membranes. But the skeptic has a real point if the question is only who declines, not how — and Experiment 2 below is built to answer it honestly: if electromyographic features add nothing beyond grip strength plus a baseline inflammation snapshot, the mechanistic framing earns nothing practical and we will say so.

Why not just use the routine blood-count ratios and skip the sub-cohort serology? For the discovery analysis, we agree — that is precisely why the primary exposure is the cheap blood-count trajectory rather than assay-heavy immunophenotyping. But the Rugao study itself found its ratios predicted only incident, not prevalent, frailty, and blood-count ratios cannot distinguish waning antibodies from stable cellular memory, which the longitudinal studies show diverge sharply (Longitudinal Analysis of Humoral and Cellular Immune Response up to 6 Months after SARS-CoV-2 BA.5/BF.7/XBB Breakthrough Infection and BA.5/BF.7-XBB Reinfection). The sub-cohort exists to test whether our trajectory features track that divergence; if they do not, the interpretation of the main result narrows to "generic inflammation," which we would report as such. We are candid that 60 participants cannot power a new immunological discovery — the sub-cohort validates the exposure measurement, nothing more.

Why not just run a randomized controlled trial? The exposure of interest — years of accumulated inflammation and immune history — cannot be assigned, and the exercise literature shows the relevant training adaptations unfold over months to years. Where randomization is infeasible or unethical, retrospective designs are the accepted standard: the delayed-intubation question could only be answered observationally (Timing to intubation COVID-19 patients: can we put it off until tomorrow?), and the antiviral effectiveness question required 497,499 patients across 29 record systems, far beyond trial feasibility (Real-World Effectiveness of Nirmatrelvir in Protecting Long COVID for Outpatient Adult Patients – A Large-Scale Observational Cohort Study from the RECOVER Initiative). The real threat is residual confounding — undocumented fitness and health-seeking behavior, exactly the limitation the retrospective literature names — and we counter it with weighting, negative-control outcomes, and explicit bias analysis rather than by claiming equivalence to a trial.

Proposed Experiments and Decisive Checks

Experiment 1 tests the association: across the 300-person cohort, inflammation-trajectory features predict the peripheral-share axis of the fatigue fingerprint after inverse-probability weighting, with a pre-specified effect size threshold and interaction test for training modality. The decisive check is a null result under adequate power — if the trajectory explains under about 5% of fingerprint variance, the core claim is false and the note ends there.

Experiment 2 tests incremental value: three nested models (gross endpoints alone; plus baseline inflammation snapshot in the Rugao specification; plus electromyographic mechanistic features) predict the fingerprint and, over an 18-month passive follow-up through the health record, incident fatigue-related encounters. The decisive metric is the area-under-the-curve increment with bootstrap intervals; a near-zero increment from the electromyographic layer would concede the skeptic's simplest alternative.

Experiment 3 tests the immune specificity: in the sub-cohort, does humoral waning (low current neutralizing titer despite documented vaccination or infection) correlate with the peripheral axis more than the central axis, consistent with the mechanism argument? This is deliberately framed as a confirmatory check on 60 people, with a clear statement that it is underpowered for anything exploratory. Together, the three experiments either establish a mechanistic, record-based predictor of how muscle fails, or cleanly delimit the cheaper tools that already suffice.