When a hospital, school, or firm rolls out a new software tool, the usual evidence of success is a survey that asks whether people like it, or a raw count of how often it gets used. Both mislead. Surveys routinely report enthusiasm for tools that nobody ends up using. Raw usage counts rise and fall with the workload itself: more patients on a Monday, more invoices at month end, more lessons during term time. A usage spike can therefore look like adoption, and a holiday lull can look like abandonment. This proposal treats usage logs as the product of two separate things. One is how many opportunities to use the tool arise, which follows the institution's calendar. The other is how likely a person is to take each opportunity, which is the actual adoption signal. We estimate the two jointly and add one component that neither adoption research nor time series research currently measures: how a person's propensity to use the tool changes when they are busiest.
The core claim is that this "load elasticity of adoption" is the quantity that separates tools with real workflow value from tools that are pleasant but disposable. A tool with real value is used at least as often per opportunity when work is heavy. A tool that adds friction is dropped exactly when pressure peaks. The adoption literature has seen this failure qualitatively, for instance in clinicians reporting information overload from a chat interface in high-tempo care. It has never measured it from logs. The decomposition literature has the machinery to measure it but has only been applied to power grids, traffic, and weather.
What each direction offers and where the gap is
The adoption literature supplies the motivating evidence. In a rural MDR-TB program, health workers rated an mHealth app as highly acceptable yet barely used it (A Pilot Study of an mHealth Application for Healthcare Workers). In a structural model of 164 endoscopists, general attitude had no significant direct effect on acceptance (p = 0.266); trust and specific beliefs carried the whole effect and explained 47% of its variance (Modeling the influence of attitudes, trust, and beliefs on endoscopists' acceptance). In a randomized trial of GutGPT, an LLM tool for gastrointestinal bleeding, the tool raised perceived ease of use (median change 0.6 vs. 0.3) but did not raise intention to adopt, and decision accuracy stayed near 92% in both arms. Trainees blamed verbosity under time pressure and the need to leave the health record system (Usability and adoption in a randomized trial of GutGPT). These three papers motivate the proposal: they show that stated attitude and measured usability do not predict use. All three are cross-sectional or vignette-based, however, and none observes behavior over weeks of real workload.
The time series side supplies the methods. N-BEATS restricts parts of a deep forecaster to polynomial (trend) and Fourier (seasonal) bases and still beat the M4 competition winner by 3% (N-BEATS). CoST learns separate trend and seasonal representations with contrastive losses and reports a 21.3% average MSE improvement on multivariate benchmarks (CoST). Tonekaboni et al. separate static per-series traits from time-varying states. They show that without their counterfactual regularizer the local encoder absorbs the global traits, and downstream AUPRC falls from 0.365 to 0.238 on PhysioNet (Decoupling Local and Global Representations of Time Series). We reuse ideas from all three. None of them models a signal that is the product of an exogenous demand process and a behavioral choice, and that product structure is the whole point here.
The closest existing bridges each miss one piece. Du et al. model repeated use of competing products (Bitly vs. TinyURL) as multivariate Hawkes processes, a class of event models in which past events raise or lower the rate of future ones (Modeling Adoption and Usage of Competing Products). Their model captures self-excitation, social influence, and competition, but its fixed exponential kernels cannot separate calendar rhythms or workload from learning. Zhang et al. simulate forced smart-meter adoption, show that user learning reshapes daily load curves, and identify proficiency rather than installation as the bottleneck (Modelling the Effects of User Learning on Forced Innovation Diffusion). Theirs is a forward simulation with a hand-set learning curve; we solve the inverse problem from noisy logs. The ETASC model fits country-level S-curves and treats all high-frequency variation as noise (A Cognitive Model for Technology Adoption). We argue that this high-frequency variation is where the information about workflow fit lives.
Core model
For user $i$ at time $t$, let $\lambda^{opp}_i(t)$ be the rate at which tasks arise where the tool could be used. Examples are patients seen, tickets opened, or links shared. These rates are observed from the host system's logs, independently of the tool. Let $p_i(t)$ be the probability that an opportunity is taken. Observed tool use is then a thinned version of the opportunity stream:
$$ \lambda^{use}_i(t) = \lambda^{opp}_i(t)\, p_i(t), \qquad \operatorname{logit} p_i(t) = \tau_i(t) + \beta_i(t)\, z_i(t) + s_i(t) + \varepsilon_i(t). $$
In the second equation, $\tau_i(t)$ is a slow adoption trend: learning, habit, or abandonment. $z_i(t)$ is standardized concurrent workload, for example the opportunity rate relative to the user's own median. $\beta_i(t)$ is the load elasticity, and it is allowed to drift slowly as the user gains experience. $s_i(t)$ captures any residual calendar pattern in propensity itself, such as lower use on night shifts at equal load. $\varepsilon_i(t)$ holds transient friction shocks, such as outages or interface changes. The opportunity rate carries its own seasonal structure and is modeled separately, so institutional rhythms never enter the adoption trend.
Most parts are borrowed. $\tau_i$ uses a low-degree polynomial or monotone-spline basis in the spirit of N-BEATS. Seasonal terms use Fourier bases. Users share a hierarchical prior, and static user traits such as role and seniority enter through a global encoder regularized as in Tonekaboni et al., which keeps slow-moving workload differences from leaking into trait estimates. The residual term uses a heavy-tailed likelihood, because the file notes that decomposition theory built on Gaussian noise breaks under bursty data (CMoS). What is new is not the neural parts. It is the product structure and the estimand $\beta_i$, which turns the qualitative "overload in high-tempo care" finding into a per-user, per-tool number with an uncertainty interval.
Research Program
The work has three stages. First, we build a synthetic benchmark of planted adoption trajectories. It combines realistic shift and weekly workload rhythms with known trends, known load elasticities (negative, zero, and positive), and heavy-tailed friction shocks. It follows the planted-pattern strategy Kabra et al. used to validate routine mining (Time Series Clustering for Human Behavior Pattern Mining). This benchmark tests whether each method recovers the adoption trend and the elasticity rather than just fitting the counts.
Second, we apply the model to the public Twitter competing-product telemetry from Du et al. There, "opportunity" is any link-shortening or retweet event, "use" is choosing a particular service or convention, and load is the user's current posting rate. This is the only real adoption log with a natural opportunity denominator available in the file.
Third, and most important, we analyze an in-situ deployment. A clinical decision-support tool embedded in a health record system, logged for at least three months alongside census and admission counts, gives both the opportunity stream and the workload signal. Brief trust and intention surveys at weeks 0, 4, and 12 let us test whether log-derived quantities track the mediators identified by Schulz et al. This dataset does not exist in the file and must be collected with a partner site. We state that as the main practical risk.
Baselines
The naive reference point is the raw weekly usage count per user, summarized by its linear slope over the first four weeks and forecast with a seasonal-naive rule that repeats last week's value. This is what most deployment dashboards report. Any method must beat it at forecasting held-out usage and at predicting abandonment by week 12.
The strongest simple competitor is the usage ratio: tool uses divided by opportunities, smoothed per user with a local regression. It removes workload seasonality at almost no cost. It is fitted on the same logs with its smoothing bandwidth tuned on the same validation weeks. It is also the baseline most likely to match us, so we treat it as the main comparison.
The strongest structured decomposition applied the conventional way is N-BEATS in its interpretable configuration, and CoST through the public salesforce/CoST code, both run on raw usage counts. They receive the same input windows, the same GPU budget, and the same number of hyperparameter trials as our model. The strongest existing adoption-specific model is the multivariate Hawkes process of Du et al., refit on each dataset with its convex estimator. Du et al. report that it beats Poisson, Weibull, and recency baselines in test log-likelihood, but the file gives no numeric value, so for the Twitter data we record "no reported result found on Lacuna for held-out log-likelihood of the Du et al. Hawkes model" and rerun it ourselves. No paper in the file reports any number for decomposing adoption telemetry, so no published figure exists to beat on this task.
The psychometric baseline is the week-0 behavioral-intention score, and the week-4 trust score where collected. This is the standard adoption-science predictor.
Two ablations isolate the key ingredients. Removing the opportunity offset tests whether the product structure matters. Fixing $\beta_i = 0$ tests whether load coupling adds anything beyond the ratio. We count our model as beating a baseline only if it improves held-out per-event log-likelihood and week-12 abandonment AUROC with bootstrap intervals that exclude zero, and, on synthetic data, recovers planted trends and elasticities with clearly higher correlation.
Why not just…?
Why not just divide usage by workload? This is the real threat. If people's propensity to use the tool does not depend on how busy they are, then the ratio and our $\tau_i$ are the same quantity, and the ratio is simpler. The ablation with $\beta_i = 0$ settles the question directly. If fitted elasticities are indistinguishable from zero across users and tools, and the $\beta_i = 0$ model matches the full model on abandonment prediction, we will report that the ratio suffices and that the only contribution is the evidence for that. The GutGPT trainees' complaints about overload under time pressure suggest otherwise. The ratio also silently averages low-load and high-load weeks, so a tool that is dropped under pressure looks moderately adopted, and the synthetic negative-elasticity condition will show whether that averaging hides planted abandonment.
Why not just run repeated surveys? Surveys measure what the Schulz et al. and GutGPT studies found to be weak predictors of use: attitude and ease of use. The MDR-TB pilot shows high acceptability coexisting with minimal real use. Surveys also cannot be given daily in acute workflows. We do not discard them. We include survey intention and trust as a baseline and test whether log-derived trend and elasticity predict week-12 use beyond them. If surveys win, that is a publishable and useful negative result.
Why not just apply an off-the-shelf decomposition such as N-BEATS or CoST to raw counts? Those methods separate trend from seasonality within a single signal. In usage logs, the seasonal part of the count is mostly workload, not behavior, and a period in which adoption and workload rise together will be split arbitrarily between the two. The synthetic "seasonal surge with no adoption change" condition is built to expose this. We expect raw-count decompositions to report a spurious adoption trend there and our model not to. This is a claim the experiment could refute.
Why not just use a Hawkes process? Du et al. show that Hawkes models capture recency and competition well, and we keep them as a baseline. Their fixed exponential kernels contain no calendar or workload term, so any load-driven drop in use would be absorbed into lower baseline rates. On the Twitter data we will also fit a Hawkes model with our opportunity offset. If that hybrid matches the full model, the right conclusion is that the offset matters and the neural decomposition does not, and we will say so.
Proposed Experiments
On the synthetic benchmark, we vary the sign and size of the planted load elasticity, the strength of workload seasonality, and the tail weight of friction shocks. We report correlation between recovered and true trends, error and interval coverage for $\beta_i$, and the rate of false "abandonment" calls during planted seasonal troughs.
On the Twitter data, we compare held-out per-event log-likelihood and next-week choice accuracy against the Hawkes, ratio, and raw-decomposition baselines. We also test whether users with negative fitted elasticity toward one service switch to a competitor during their own posting bursts.
On the clinical deployment, the decisive check has two parts. First, the log-derived trend and elasticity from weeks 1 to 4 should predict whether a clinician is still using the tool at week 12, with higher AUROC than week-0 intention and than the ratio slope. Second, fitted $\beta_i$ should correlate with week-4 trust scores more strongly than raw usage volume does, which would give the first log-based handle on the trust mediator Schulz et al. identified.
A pre-registered falsification condition frames the whole project. If elasticities are near zero everywhere, and the ratio baseline matches the full model on all three datasets, the central claim fails. The proposal would then reduce to a recommendation to always normalize usage by opportunities, which is still worth publishing but is not the result we are betting on.
References
- A Pilot Study of an mHealth Application for Healthcare Workers: Poor Uptake Despite High Reported Acceptability at a Rural South African Community-Based MDR-TB Treatment Program
- Modeling the influence of attitudes, trust, and beliefs on endoscopists’ acceptance of artificial intelligence applications in medical practice
- Usability and adoption in a randomized trial of GutGPT a GenAI tool for gastrointestinal bleeding
- N-BEATS: Neural basis expansion analysis for interpretable time series forecasting
- CoST: Contrastive Learning of Disentangled Seasonal-Trend Representations for Time Series Forecasting
- Decoupling Local and Global Representations of Time Series
- Modeling Adoption and Usage of Competing Products
- Modelling the Effects of User Learning on Forced Innovation Diffusion
- A Cognitive Model for Technology Adoption
- CMoS: Rethinking Time Series Prediction Through the Lens of Chunk-wise Spatial Correlations
- Time Series Clustering for Human Behavior Pattern Mining