CosinorAge: Digital Biological Age Estimation

• CosinorAge is a digital biomarker framework that computes a mortality-based biological age from wearable-recorded rest–activity and sleep metrics.
• The framework standardizes minute-level ENMO across devices and employs cosinor analysis to extract rhythmic and physical activity features.
• It integrates an end-to-end open-source pipeline from data ingestion and preprocessing to feature computation and risk-calibrated age estimation.

Searching arXiv for the cited papers and related context. CosinorAge is a digital biomarker framework and software ecosystem for estimating biological age from the rest–activity rhythms, physical activity, and sleep patterns captured by wearable devices and smartwatches. It consists of an open-source Python package plus a web-based “CosinorAge Calculator” that expose the same analytical pipeline, with an end-to-end workflow from raw data ingestion and preprocessing to feature computation and biological age estimation. The framework is explicitly positioned as a mortality-based, “second-order clock” derived from wearable-derived circadian, physical activity, and sleep metrics rather than a model trained only to predict chronological age (Shim et al., 1 Sep 2025).

1. Conceptual basis and aging interpretation

CosinorAge is built on the premise that patterns of movement and sleep across the 24-hour day contain information about underlying physiological state and aging trajectory. In this formulation, circadian rhythms orchestrate metabolic, immune, endocrine, and other regulatory systems, and the daily cycle of rest and activity—including sleep timing, sleep quality, and daytime activity—reflects the integrity of those rhythms. Lower rhythm amplitude, irregular activity timing, chronically low activity levels, and poor or irregular sleep are described as having been associated with higher risks of chronic disease, mortality, systemic inflammation, and accelerated biological aging (Shim et al., 1 Sep 2025).

A central distinction in the framework is between chronological age and biological age. Chronological age is time since birth. Biological age is intended to quantify the state of the body’s systems relative to typical aging trajectories, such that physiology may appear older or younger than calendar age. CosinorAge is explicitly described as a “second-order clock” because it is derived from mortality risk and is designed to capture healthspan-relevant processes and risk, in analogy to epigenetic clocks but using digital behavioral data from wearables (Shim et al., 1 Sep 2025).

The choice of wearable-derived data is integral to that framing. Smartwatches and fitness trackers continuously record movement via accelerometers and, in many devices, sleep–wake patterns inferred from movement and/or other sensors. These streams enable continuous, high-resolution monitoring of rest–activity rhythms at scale, in real-life environments, non-invasively and cheaply. The framework therefore aims to transform routine step, activity, and sleep tracking into a reproducible estimate of biological age and rhythm health.

A common misconception is to treat CosinorAge as a direct clinical judgment. The framework instead defines it as a probabilistic digital biomarker derived from behaviorally measured rhythms, physical activity, and sleep. The source description explicitly states that it should not be used as a standalone diagnostic and should be interpreted alongside other health indicators.

2. Data model, supported inputs, and software architecture

The core data type in CosinorAge is accelerometry-based activity intensity standardized as ENMO (Euclidean Norm Minus One) at minute-level resolution. ENMO is described as a widely used summary of raw triaxial acceleration that approximates movement intensity in milli-gravity units. The pipeline standardizes different input sources to minute-level ENMO time series, which then serve as the common basis for circadian rhythm analysis, physical activity metrics, and sleep-related metrics (Shim et al., 1 Sep 2025).

The framework supports multiple data sources. Research-grade wrist accelerometer data from UK Biobank are processed through UKBDataHandler, NHANES accelerometry data through NHANESDataHandler, Samsung Galaxy Watch data through GalaxyDataHandler, and bring-your-own-data CSV inputs through GenericDataHandler. The BYOD path requires time-stamped activity or acceleration measurements convertible to minute-level ENMO. This design means that many other devices can be supported if their data can be transformed to the required format.

Across sources, handlers normalize sampling rate and resolution, resample or aggregate to minute-level ENMO, standardize timestamps, and align data on a continuous time axis. They also distinguish valid wear time from non-wear periods, with non-wear intervals excluded from feature computation and cosinor fits. The detailed preprocessing logic, including exact thresholds or algorithms, is described as documented on GitHub rather than exhaustively specified in the software paper.

The Python framework is organized into three core modules.

Module	Role	Main output
DataHandler	Data ingestion and preprocessing	Standardized minute-level ENMO
WearableFeatures	Feature extraction	Circadian, activity, and sleep features
CosinorAge	Biological age estimation	Mortality-based biological age

Within WearableFeatures, the package distinguishes WearableFeatures for individual-level analysis and BulkWearableFeatures for cohort-level batch processing. The CosinorAge module then applies pre-trained model coefficients to the derived features and returns an estimated biological age in years. The model supports unisex, female-specific, and male-specific coefficient sets, with sex optionally supplied to select the corresponding model (Shim et al., 1 Sep 2025).

3. Cosinor analysis and rhythmic parameterization

A central analytic component of the framework is cosinor analysis of rest–activity rhythms. The generic model is given as

$$y(t) = M + A \cos(\omega t + \phi) + \varepsilon(t),$$

where $y(t)$ is observed activity such as ENMO at time $t$, $M$ is the MESOR, $A$ is the amplitude, $\phi$ is the acrophase, $\omega$ is the angular frequency, typically $2\pi/24$ for daily rhythms, and $\varepsilon(t)$ is residual error (Shim et al., 1 Sep 2025).

Within this formulation, MESOR is the rhythm-adjusted mean activity level, amplitude quantifies the strength of day–night contrast, and acrophase encodes the timing of peak activity. The source description states that these parameters are sensitive to rhythm robustness and misalignment, and are linked to health and aging. CosinorAge uses these parameters together with nonparametric circadian metrics rather than treating the cosine fit as a standalone summary.

The broader cosinor literature included in the source material highlights a methodological issue relevant to rhythmic biomarker analysis: harmonic underspecification. In a standard trigonometric regression of order $K$, fitting too few harmonics relative to the true oscillatory structure can bias estimated coefficients in some model classes. Under equispaced sampling and generalized gamma assumptions, least-squares trigonometric regression on $y(t)$0, described there as log-normal trigonometric regression, yields unbiased harmonic parameter estimates for the included harmonics even when the number of oscillation harmonics is underspecified, whereas GLMs with log link require correct specification to produce unbiased harmonic parameter estimates (Gorczyca, 16 Jan 2026).

This does not alter the software description of CosinorAge itself, which is framed around standard cosinor-style rhythmic summaries, but it clarifies a methodological boundary around cosinor-based inference. A plausible implication is that the robustness of estimated MESOR, amplitude, and acrophase depends not only on the chosen rhythm model but also on distributional assumptions and harmonic specification. That issue is particularly relevant whenever rhythmic biomarkers are positive, skewed, or potentially multi-harmonic (Gorczyca, 16 Jan 2026).

4. Feature engineering across circadian, activity, and sleep domains

The WearableFeatures module computes three broad families of features from minute-level ENMO: parametric and nonparametric circadian rhythm metrics, physical activity metrics, and sleep metrics. These features constitute the covariates for the mortality-based biological age model (Shim et al., 1 Sep 2025).

The circadian feature set includes MESOR, cosinor amplitude, acrophase, M10, L5, relative amplitude, interdaily stability, and intradaily variability. M10 is the average activity during the 10 consecutive hours with highest ENMO values, while L5 is the average activity during the 5 consecutive hours with lowest ENMO values, typically representing nighttime rest. Relative amplitude is described as a composite index combining M10 and L5:

$y(t)$1

Interdaily stability quantifies regularity of the 24-hour pattern across days, and intradaily variability measures fragmentation of activity within the day. Together these metrics characterize rhythm strength, timing, regularity, and fragmentation.

The physical activity domain is based on ENMO thresholds, documented as device-specific or dataset-specific on GitHub. From these thresholds, the framework computes light physical activity, moderate physical activity, vigorous physical activity, and sedentary duration. The description characterizes these quantities as behavioral dosage measures related to cardiometabolic health and mortality risk.

The sleep-related feature set is derived from accelerometry-based behavioral patterns rather than direct sleep stage inputs. It includes total sleep time, wake after sleep onset, percent time asleep, number of waking bouts, and sleep onset latency. These metrics reflect sleep quantity, continuity, and fragmentation.

Feature computation is standardized around minute-level ENMO as a common input. For multi-day recordings, features are typically averaged or otherwise aggregated across days to produce more stable estimates of trait-like rhythms rather than single-day fluctuations. The description notes that, while exact scaling procedures are not listed, features are standardized in practice when fitting and applying the proportional hazards models, with the same scaling applied when using the open weights (Shim et al., 1 Sep 2025).

5. Mortality-based age modeling and relation to other aging frameworks

CosinorAge maps wearable-derived features to mortality risk and then to a biological age estimate via a proportional hazards survival model. The model is described as akin to a Cox model:

$y(t)$2

where $y(t)$3 is hazard of death at time $y(t)$4 conditional on feature vector $y(t)$5, $y(t)$6 is baseline hazard, $y(t)$7 is the vector of coefficients, and $y(t)$8 is the vector of circadian, physical activity, and sleep features (Shim et al., 1 Sep 2025).

The framework then defines a risk-based biological age. The source description does not provide the explicit conversion formula, but it states that mortality-based clocks define biological age as the age for which an individual’s estimated hazard equals the average hazard at that age in the training cohort. In abstract form, the risk score is

$y(t)$9

and CosinorAge is obtained by mapping $t$0 to an equivalent age through a cohort-derived age–hazard relationship. The resulting output is a biological age estimate in years. The source gives an illustrative example in which minute-level ENMO from a Samsung Galaxy Watch over 7 days for a 45-year-old female yielded a predicted biological age of 49.0 years.

The same source also presents the conceptual age-acceleration quantity

$t$1

with positive values interpreted as older-appearing physiology and negative values as younger-appearing physiology, conditional on the mortality-based calibration.

Within the broader landscape of aging models, CosinorAge differs from frameworks that infer aging structure from cross-sectional multivariate biomarker data without explicit circadian feature engineering. One such latent-variable approach represents each individual’s features as a nonlinear function of a low-dimensional, linearly evolving latent state and proves identifiability from cross-sectional data under order-isomorphic constraints (Pierson et al., 2018). That model targets multidimensional rates of aging from cross-sectional observations rather than digital circadian and behavioral features, but it addresses a related conceptual question: how latent temporal information can be inferred from non-longitudinal data (Pierson et al., 2018).

This comparison clarifies what CosinorAge is and is not. It is not a generic latent aging model and not a direct chronological-age regressor. It is a specific mortality-based aging clock built from wearable-derived rest–activity, physical activity, and sleep features. A plausible implication is that its primary interpretive axis is risk-calibrated behavioral aging rather than a fully multidimensional latent decomposition of aging processes.

6. Web platform, reproducibility, limitations, and applications

The web-based CosinorAge Calculator at www.cosinorage.app exposes the same analytical pipeline as the Python package. Users upload CSV files containing time-stamped activity or accelerometry data, preview file structure, configure device type, timestamp format, time zone, and analysis parameters, and then invoke server-side processing through the Python package: DataHandler to produce minute-level ENMO, WearableFeatures for feature computation, and CosinorAge for biological age estimation (Shim et al., 1 Sep 2025).

Outputs include report-style visualizations and summaries: individual-level rhythm plots, summary tables of circadian, activity, and sleep features, and the biological age estimate. In multi-individual mode, the application provides descriptive statistics across the cohort, a feature correlation matrix, and visual summaries such as feature distributions. The interface is organized into Home, Documentation, Calculator, and About sections.

Reproducibility and transparency are central design goals. The framework is explicitly intended to address the lack of unified open-source pipelines for jointly analyzing rest–activity rhythmicity, physical activity, and sleep and linking them to health outcomes. The software description contrasts this with tools such as pyActigraphy, actipy, CosinorPy, scikit-digital-health, and GGIR, which are described as focusing on specific domains or being tied to R and not connecting metrics to health or aging outcomes. CosinorAge is presented as integrating circadian, activity, and sleep analyses, attaching a mortality-based biological age model, and distributing the model coefficients as open weights (Shim et al., 1 Sep 2025).

Several limitations are stated directly. Different devices may differ in accelerometer sensitivity, placement, and data quality; ENMO standardization helps, but cross-device comparability studies remain important. Models trained on UK Biobank may not generalize perfectly across populations that differ in age, ethnicity, or lifestyle. Biological age estimates are derived from behavior and rhythms, not from direct physiological measures. For these reasons, the framework is described as a digital biomarker of aging and resilience rather than a definitive clinical judgment (Shim et al., 1 Sep 2025).

The applications listed in the source include personalized health monitoring, aging and resilience research, intervention studies, and possible future clinical and digital health integration. Individuals may track changes in CosinorAge over time in relation to sleep or activity interventions. Researchers may use it to quantify rest–activity patterns in large cohorts and study associations with aging trajectories and healthspan outcomes. The source further suggests future directions involving heart rate, HRV, temperature, integration with epigenetic clocks, lab biomarkers, clinical records, and more advanced modeling techniques, while emphasizing that the current implementation is proportional hazards-based and accelerometry-centered (Shim et al., 1 Sep 2025).