Separating Tracker Novelty from Physiology in Wearable Biomarker Models

When someone starts wearing a new fitness tracker, they often walk more, sleep more carefully and check their heart rate more often for the first few weeks, and then drift back toward their usual habits. This is the novelty effect: a temporary change in behavior caused by the newness of a tool rather than by anything the tool does. It is well documented in education and product testing. Models that predict clinical blood markers from wearable data treat these first weeks as ordinary physiology. The clearest example predicts insulin resistance from resting heart rate, steps and heart rate variability averaged over 7- to 90-day windows (Insulin resistance prediction from wearables and routine blood biomarkers). If the early windows are inflated by novelty, a 14-day step average describes the person's honeymoon with the device, not their metabolic lifestyle.

This note proposes to model that honeymoon explicitly. Each person's daily wearable signal is treated as a stable personal level plus a person-specific transient that decays after the device is first worn, plus weekly rhythm and noise. The claim is that estimating and removing the transient does two useful things. It lets a model infer a person's long-run level, and so their blood-marker risk, from only one or two weeks of data. It also gives the first objective, continuous, per-person measurement of how fast a novelty effect fades.

Core Claim

The wearable features that feed biomarker prediction are biased in a structured, time-since-onset-dependent way during the first weeks of device use. Because that bias follows a common shape, a hierarchical model that pools the decay shape across people can recover each person's stable level from a short early record. A raw short-window average, which averages the transient in, cannot do this. Neither can discarding an initial "burn-in" period, which throws away most of a short record. If the claim holds, insulin-resistance prediction from 14 days of transient-corrected data should approach the accuracy of 90-day windows. The claim also has a falsifiable precondition. The transient should appear as a function of days since device onset, separable from calendar season, and it should be absent or much smaller in people who already owned a tracker.

Why This Gap Exists

The two literatures have not met. On the novelty side, the evidence is behavioral and subjective. The move from paper to tablet textbooks raised a coded engagement indicator from 0.56 to 0.87, and the authors flagged novelty as an uncontrolled confound they could not separate from real benefit (The Comparison of Paper Textbook Class and Electronic Textbook Class in Technology Rich Classroom). Meta-analyses show gains fading after about ten weeks in flipped classrooms (Effects of flipped language classrooms on learning outcomes in higher education: A Bayesian meta-analysis) and vanishing in long-term AI-blended learning (Unlocking the potential of artificial intelligence in improving learning achievement in blended learning: a meta-analysis). These papers motivate the proposal by showing that early windows mislead. None of them measures the decay in any continuous objective signal.

The closest methodological work is Estimating Effects of Long-Term Treatments. It corrects novelty-distorted A/B tests by extrapolating long-term effects from early "surrogate" outcomes using linear surrogate indices, and it reduced bias by 59.8% on average on WeChat Search experiments. We borrow its core idea that early trajectories can forecast late outcomes. Our setting differs in three ways. Its target is a population treatment effect, while ours is an individual's stable physiological level. Its surrogacy assumption is untestable, whereas ours can be checked directly against held-out late-window data. Its model is a pooled linear regression, while ours fits a per-person decay curve.

On the biomarker side, trajectory work shows that the shape of a series carries information a snapshot does not. Examples include serum albumin in myocardial infarction (Association between albumin changes and prognosis in older patients with acute myocardial infarction) and hsCRP trajectory classes predicting metabolic syndrome (Association Between High-Sensitivity C-Reactive Protein Trajectories and the Incidence of Metabolic Syndrome:A Retrospective Cohort Study). Both motivate modeling trajectories rather than averages. Neither considers that the act of starting measurement can itself move the signal.

The method we reuse most directly comes from A Bayesian Approach to Modeling Variance of Intensive Longitudinal Biomarker Data as a Predictor of Health Outcomes. That paper jointly models a wearable heart-rate series, with spline mean and autoregressive residuals, together with a health outcome. In simulation, the separate two-stage approach dropped to 11.5% coverage while the joint model held 94.5%. This result constrains our design: we fit the transient and the outcome jointly rather than plugging estimated levels into a downstream regressor. That work covered minutes of lab stress, not weeks of habituation, so extending it to a multi-week decay component is our contribution.

Research Program

For person $i$ on day $t$ since first wearing the device, we model a wearable feature $y_{it}$ (daily steps, resting heart rate, or heart rate variability) as

$$y_{it} = \mu_i + a_i e^{-t/\tau_i} + w_{d(t)} + \epsilon_{it},$$

where $\mu_i$ is the person's stable level, $a_i$ is the size of their novelty transient, $\tau_i$ is its decay time in days, $w_{d(t)}$ is a day-of-week effect, and $\epsilon_{it}$ is an autoregressive residual. The parameters $a_i$ and $\tau_i$ get population-level priors, so a person with only 14 days borrows the typical decay shape from everyone else. This pooling is what makes short-record extrapolation possible.

The exponential form is a starting hypothesis, chosen because habituation is classically modeled as a first-order decay. We will compare it against a spline in days-since-onset shared across people, with a person-level amplitude. The stable levels $\mu_i$ for all features then enter a probit or linear outcome model for HOMA-IR, the standard insulin-resistance index computed from fasting glucose and insulin. Everything is estimated jointly by Markov chain Monte Carlo, following the variance-modeling paper.

The primary data source is the WEAR-ME cohort: 1,165 participants with Fitbit or Pixel Watch streams and blood panels, plus a 72-person validation cohort. Access to daily-resolution data with device-onset dates is the main feasibility risk. If WEAR-ME cannot be obtained, the protocol applies unchanged to any cohort that records device start dates and a blood panel, but then we would have no published baseline to compare against.

Attrition is a known issue: the analyzed cohort was about a quarter of enrollees. We therefore define targets only among people with 90-day coverage. Separately, we report whether early transient parameters predict dropout.

Baselines

The naive reference is the raw k-day average fed to the published XGBoost pipeline, for k equal to 7, 14 and 30. This is how the insulin-resistance paper builds its short-window features. The strongest existing result on this question is that paper's best multimodal model, which reports $R^2$ of 0.50 for HOMA-IR regression and AUROC of 0.80 for HOMA-IR at or above 2.9 (Insulin resistance prediction from wearables and routine blood biomarkers). The file does not give per-window numbers, so there is no reported result found on Lacuna for 14-day-only performance. We will reproduce that per-window breakdown ourselves before comparing.

The burn-in baseline discards the first b days, with b tuned on validation data, and averages the rest. The index-aware baseline gives XGBoost per-day values and days-since-onset as features. A strong learned baseline is a sequence model trained to predict a person's day 61–90 mean directly from their first k days. We use the masked autoencoder family the WEAR-ME authors already applied, with a regression head.

The key ablation is our own hierarchical model with $a_i$ fixed at zero. This keeps pooling, weekly effects and joint estimation, and removes only the transient. It separates the value of modeling novelty from the value of Bayesian shrinkage.

All methods use identical participants, splits, blood-panel covariates and demographic features. Each gets the same hyperparameter search budget, measured in validation evaluations. Beating the field means two things. First, lower error forecasting the day 61–90 mean from 14 days than every baseline, with a paired bootstrap interval excluding zero. Second, 14-day HOMA-IR AUROC closer to the 90-day figure than any baseline achieves.

Why not just…?

Why not just discard the first few weeks? Burn-in is the obvious fix, and at 90 days it will probably work about as well as our model. We expect to concede that regime. Its failure is at short horizons. With a 14-day record and a transient lasting two or three weeks, burn-in leaves little or no data, which forces the user to wait. Given heavy attrition, many people never reach the end of the burn-in. The comparison at k equal to 7, 14 and 30 days settles where, if anywhere, explicit modeling pays off.

Why not just use the 90-day window, since it already exists? For risk screening, a model that needs three months of wear loses exactly the people who stop wearing the device. The value of the proposal is earlier prediction, so the 90-day model is the target to match with less data, not a competitor to beat.

Why not just let a flexible model learn the transient from days-since-onset or raw sequences? This is the real threat. A sequence model with enough data could learn per-person extrapolation implicitly. With about a thousand people, we expect the pooled parametric decay to be more sample-efficient. We will test this by subsampling training sets from 200 to the full cohort. If the sequence model matches at full size, the contribution narrows to the measured novelty curve and its uncertainty estimates, and we will report it that way. The index-aware XGBoost is weaker by construction: it learns a population curve, not a person-specific amplitude and decay.

Decisive Checks

The first check comes before any prediction. Plot population wearable features against days since onset, controlling for calendar date by exploiting staggered enrollment. A transient should appear in steps and possibly heart rate variability. If the cohort records prior tracker ownership, the transient should be clearly larger in first-time owners. If no transient survives the calendar controls, the proposal's premise fails, and we report that negative finding.

The second check is out-of-sample forecasting of each person's late-window mean from early data, compared across all baselines.

The third check is HOMA-IR prediction with k-day features, including the 72-person validation cohort. We also test whether transient-corrected levels reduce the gap between early and late HOMA-IR predictions within the same person.

The last analysis feeds back to the novelty literature. We report the distribution of per-person decay times $\tau_i$ and ask whether it resembles the several-week horizons reported in educational settings. We also ask whether large transients predict dropout. This would supply the objective, continuous habituation measure that the classroom studies lacked.