Public methods specification

Model specification.
Evidence boundaries.

This page separates published methods and supporting evidence from BioMIR-specific implementation choices. It documents what is computed, how the layers relate, what the quality indicators mean, and which claims require independent validation.

Research & validation status

Interpretation boundary

Specification is not validation.

Published methods can provide a scientific foundation without validating a particular software implementation, biomarker panel, coefficient set, composite, classification system, or user population.

BioMIR therefore distinguishes four questions: what is measured, what the literature supports, what the software specifies, and what has been empirically validated. Evidence for component biomarkers does not automatically validate the composite.

Terminology in BioMIR

BioMIR uses “biomarker” as a broad product umbrella connecting reported dietary and behavioral inputs, observed physiological and clinical measurements, and modeled indicators. The FDA–NIH BEST resource also includes physiological measures such as blood pressure in biomarker terminology. BioMIR identifies each item as a reported input, an observed measurement, or a derived estimate so its source and role are clear.

Subclinical Analysis is BioMIR’s label for seven synthetic scenarios spanning a stable baseline and selected exposure or response patterns. The cardiometabolic scenario is informed by the AHA cardiovascular-kidney-metabolic (CKM) staging framework, but it is not a formal CKM stage, diagnosis, or clinical risk estimate. BioMIR models selected inputs and does not apply the AHA framework’s full kidney, cardiovascular, or risk-calculator criteria. See the 2026 AHA/ACC/ADA/ASN CKM guideline for current clinical guidance. The label names a software scenario set; it does not assert disease status, symptoms, or clinical validation.

Canonical temporal architecture

Behavioral → Functional →
Cardiometabolic → Clinical Long View.

The ordering organizes characteristic observation cadence and biological role. It is an interpretive architecture—not a deterministic causal chain.

01 / Behavioral

Modifiable exposures and routines

The Behavioral domain uses Steps, Dietary Energy, Dietary Carbohydrates, Time in Daylight, Alcohol, and Dietary Sodium; Meditation is not active. Its configured responses form a signed, bounded modifier on a provisional age-equivalent scale. The model applies BioMIR-directed favorable adjustments to its lowest Dietary Energy band and low carbohydrate-share region; these adjustments are model-design rules, not mortality-derived protective effects. Other response regions retain their configured source-informed functions. Scoreable inputs are renormalized over the inputs available for that day; missing entries are not imputed as zero. Alcohol may display a default of zero, but an unreported value is not scored, and a daily report does not establish habitual exposure.

Behavior can change within hours. The domain represents recorded exposure and routine; it does not imply that one logged behavior causes a same-day physiological or clinical outcome.

02 / Functional

Regulation, restoration & capacity

The Functional domain integrates autonomic regulation, cardiorespiratory capacity, post-exercise response, mobility, and restorative physiology. Its six inputs are VO₂max, Resting Heart Rate, Heart Rate Variability, Cardio Recovery, Walking Speed, and Total Sleep. The provisional source-informed mappings produce an adverse-only, nonnegative modifier. Scoreable inputs use their configured weights renormalized over the inputs available for that day; missing observations are not imputed as zero. The input functions, cap limits, and conversion to age-equivalent units remain subject to qualification and independent evaluation.

Some measures vary from day to day, while aerobic fitness integrates a longer adaptation window. Wearable estimates remain subject to device and algorithm measurement error.

03 / Cardiometabolic

Metabolic, vascular & body-composition state

Systolic blood pressure, fasting blood glucose, and BMI are combined using age-regressed population coefficients to construct the Cardiometabolic Age baseline. Each marker remains independently dated. BioMIR preserves the original glucose-selection behavior: explicitly fasting clinical glucose is preferred when available, while operational fallback glucose observations can supply the modeling input without establishing laboratory-confirmed fasting status.

When a new eligible observation is unavailable, the latest eligible value can be carried forward without changing its original observation date. Data Freshness decreases as the supporting observations age.

04 / Clinical Long View

Episodic multivariate clinical-panel context

KDM and Levine PhenoAge are computed from their required clinical measurements and laboratory biomarkers when suitable dated records are available. They remain separate from Adaptive BioAge rather than being folded into the daily composite.

Their role is comparative clinical context across a longer horizon. Differences between clocks can reflect different biomarker panels, statistical objectives, collection dates, and reference assumptions.

Deterministic scientific core

Same inputs. Same model version. Same scientific result.

BioMIR uses prespecified calculation rules and scientific settings so that the same eligible measurements, reference context, and model version produce reproducible results. Population-reference values are derived from identified source populations and are interpreted separately from outcome evidence and clinical guidance.

The authoritative scientific calculation is not a self-training model: adaptive or generative components are not permitted to silently change the coefficients, reference rules, bounds, or equations used to produce a BioMIR result. Scientific-model changes belong to explicit, identifiable model versions so interpretation and validation can remain tied to the computation that generated the output.

This structure supports auditability, change control, implementation verification, and future formal review.

FDA · Key Considerations for Digitally Derived Measures in Clinical Investigations (2026) ↗

Composite construction

Three domains, one additive result.

Adaptive BioAge

Adaptive BioAge combines the Cardiometabolic Age foundation with bounded Functional and Behavioral modifiers, then compares the result with chronological age. Cardiometabolic Age is the age-bearing foundation; Functional represents adverse burden only, while Behavioral can be favorable or adverse. Scoreable inputs are renormalized over the available panel, and missing values are not substituted with zero.

Relative weights

Each configured source hazard-ratio response is evaluated in log space. Functional retains adverse response deviation; Behavioral centers source responses and signs them so favorable patterns contribute favorably. Source-response magnitudes determine the configured normalized shares. The low-end Dietary Energy and carbohydrate-share adjustments are applied as separate BioMIR-directed scoring rules; they do not change the source-derived response magnitudes or represent source-established mortality protection. For each record, scoreable inputs are combined over the available panel using the configured within-domain shares.

Asymmetric domain modifiers

After the available-panel and sex-profile bounds, the mean source log-hazard departure is translated through an age- and sex-specific population mortality-rate curve. Functional contributes nonnegative adverse burden; Behavioral is signed and may contribute in a favorable or adverse direction. This asymmetry reflects biological realism: protective behavior can buffer burden, but its modeled contribution has a limit, while adverse Functional burden can persist. Cardiometabolic Age is calculated by its separate age-regression model.

Δ-years

At evaluation, ΔAdaptive BioAge is the Adaptive BioAge estimate minus chronological age. The record-level value reconciles to the additive Cardiometabolic, Functional, and Behavioral contributions.

Model integration and attribution.

Adaptive BioAge combines the Cardiometabolic Age foundation with bounded Functional and Behavioral modifiers, then compares the result with chronological age. Cardiometabolic Age is the age-bearing foundation; Functional represents adverse burden only, while Behavioral can be favorable or adverse. Cardiometabolic Age is distinct from the clinical KDM model.

Configured response shapes. Functional uses the two-sided Total Sleep categories (<6 hours, 6–8 hours, and >8 hours), with linear or log-linear responses for the five other measures. Behavioral uses a nonlinear Steps curve, sex-specific habitual-intake categories for Alcohol, paired same-day carbohydrate percent-energy and Dietary Energy responses, ranked source responses for Dietary Energy and Sodium, and annual daylight-duration anchors. The lowest Dietary Energy band and low carbohydrate-share region receive separate BioMIR-directed favorable adjustments; these are scoring rules rather than mortality-derived effects. Sodium retains its configured ranked response. An unreported Alcohol default is not treated as a measured exposure. Scoreable inputs use their configured weights, renormalized over the inputs available for that record.

Score and age mapping. Published source responses are evaluated in log space, then passed through the configured domain scoring rules. For Behavioral, the separate low-end Dietary Energy and carbohydrate-share adjustments are BioMIR-directed and are not source mortality effects. Source-response magnitudes determine normalized within-domain shares. After panel-specific bounds, the bounded domain score is rescaled and mapped through an age- and sex-specific population mortality-rate curve. The final model remains additive. Cardiometabolic Age and the mortality-rate mapping are separate calibrations.

Public model framework and evidence status · Model equations. Proprietary numerical parameters are not part of this public description.

Functional measures and evidence-informed weighting.

The six-input Functional model uses a two-sided sleep-duration response (<6 hours, 6–8 hours, and >8 hours) and linear or log-linear source-response functions for VO₂max, resting heart rate, HRV, Walking Speed, and one-minute Cardio Recovery. The Functional transform retains adverse burden only. Scoreable inputs use their configured within-domain shares renormalized across the available panel.

Walking Speed population references use NIH Toolbox usual-pace 4-meter test means and standard deviations by age and sex; its mortality associations come from separate cohort studies. Cardio Recovery population references use Sydó et al. age-specific HRR1 means and standard deviations from a primary-prevention exercise-test cohort (ages 30–79); its mortality association comes from a separate meta-analysis. A 2025 cycle-test study reports HRR reference values by age and sex. Measured gait-speed mortality meta-analysis · Updated mixed-method dose-response meta-analysis · NIH Toolbox usual-pace 4-meter gait-speed reference values · Related UK Biobank daily-life walking measurement study · Heart-rate-recovery mortality meta-analysis · Sydó et al. age-specific HRR1 cohort values · Jou et al. age- and sex-specific cycle-test reference values, 2025 · HealthKit walking-speed definition · HealthKit one-minute recovery definition

REC: a separate action-credit record

REC is displayed alongside Adaptive BioAge and does not alter its calculation. Bounded/asymmetric model modifiers do not establish a biological credit exchange rate, response delay, or saturation law. Sustained behavior and subsequent measured FUN/CMA inputs must be evaluated separately; action credit does not establish a physiological exchange rate.

Action priority is a separate rule layer. A recent eligible Behavioral report can support a next-day suggestion, while supported seven-day persistence and trend can inform Behavioral and Functional ordering within modeled contribution bands. This ordering does not change the daily Adaptive BioAge score or treat a carried value as a new measurement; the suggested action remains tied to the native measurement and its modeled band. These are implementation rules that still require evaluation for usefulness and outcomes. Read the Today Actions explanation.

Separate clinical models

KDM & Levine PhenoAge

BioMIR presents these as periodic clinical-panel perspectives rather than interchangeable substitutes for the daily Adaptive BioAge composite.

Klemera–Doubal Method

KDM combines age-related information across a biomarker panel while accounting for biomarker-specific residual variation. Although KDM and Cardiometabolic Age both use age-related population relationships to construct age-equivalent estimates, they are not the same method: BioMIR Cardiometabolic Age is a three-marker age-regression baseline, whereas KDM is a published multivariate biological-age estimator with a distinct statistical construction.

Original method

Levine PhenoAge

The clinical PhenoAge model combines chronological age with nine laboratory biomarkers in an outcome-linked framework. It is distinct from DNAm PhenoAge, the related methylation model.

Clinical Phenotypic Age research

Results, statistics & interpretation

Support for an estimate is part of the result.

Data Confidence, Completeness & Data Freshness

Data Confidence summarizes support across 15 active Adaptive BioAge inputs: six Functional, six Behavioral, and three Cardiometabolic anchors. Each stream contributes according to its applicable same-day completeness or observation-age rule; this is not a statistical confidence interval, device-accuracy score, diagnosis, or probability that a model result is correct. ABA Trends labels the corresponding support series Completeness. Missing streams contribute zero to this support summary, but missingness is not itself a measured zero: a domain computes from scoreable inputs using the existing present-weight renormalization rule. Eight dynamic streams—HRV, resting heart rate, and the six daily Behavioral records—accrue support across the day using same-day provenance timestamps. Behavioral inputs, HRV, resting heart rate, and sleep remain attached to their own daily records; prior-day values are not carried forward. VO₂max, Walking Speed, and Cardio Recovery retain their original dates when eligible carry-forward is used and contribute linearly declining support over a 30-day operational window. BMI, systolic blood pressure, and glucose use the shared Data Freshness score. Separate Cardiometabolic and clinical-panel Data Freshness displays show observation recency by anchor or panel. Meditation is not an active scoring input; historical observations are not counted among the 15 active inputs.

Functional observation age · Data Confidence

The age of carried VO₂max, Walking Speed, and Cardio Recovery measurements reduces their support contribution from full support on the observation day to zero at 30 calendar days. This affects Data Confidence only; it does not alter the stored measurement or its modeled ΔYears contribution. The 30-day window is an operational support convention, not a biological half-life or clinical expiration.

Data Freshness · cardiometabolic anchors

Data Freshness is one model-weighted recency score derived from the independently preserved observation dates of BMI, systolic blood pressure, and fasting blood glucose. Each marker’s recency declines linearly across BioMIR’s 30-day operational horizon, and the three recencies are combined using the active CMA freshness weights. Carry-forward preserves the original date; it does not create a new measurement.

Data Freshness · clinical panel

The main KDM and clinical PhenoAge displays carry forward the latest persisted dated laboratory panel; statistical detail separately summarizes actual panels over five years. Data Freshness declines linearly from 100% on the laboratory date to 0% across BioMIR’s 180-day operational horizon while the stored clinical acceleration remains unchanged. The displayed biological age is chronological age at the selected evaluation date plus that stored acceleration; the original laboratory date is preserved. The 30- and 180-day horizons are BioMIR recency conventions, not biological half-lives, clinical expiration dates, or mandatory retesting schedules.

Statistical uncertainty

Sample count, confidence intervals, aggregation method, missingness, irregular sampling, carried observations, device error, and reference-population mismatch can all affect interpretation. No single displayed quality metric captures all of these sources of uncertainty.

External validation

The active Functional and Behavioral parameter set uses mean available marginal source log-hazard departures and a synthetic moment-matched reference distribution; it is not a fitted joint mortality model. Its uncertainty, source transport, measurement correspondence, dependence, Energy logging versus calibrated intake, and daily versus habitual alcohol exposure remain under assessment. The selected bounds do not guarantee that Cardiometabolic Age is numerically dominant for every record. Δ-years is a model-specific age-equivalent summary, not a direct measure of aging rate, lifespan, disease probability, or an intervention effect. Independent evaluation must assess calibration, longitudinal reliability, measurement correspondence and error, missingness, subgroup behavior, dependence among inputs, and incremental predictive performance against independently defined outcomes. Software checks do not establish external predictive or clinical validity, regulatory authorization, or release approval.

Period distributions and window trends

The main chart reports the selected level summary: mean, median or fitted period-end estimate (κ), where supported. The HUD links a point to its domain and reporting period. By default, ABA Trend, ABA Persistence and REC Adherence summarize the visible window. In N-mode (D, W and M), daily mean values keep the companions fixed to that window. A selected domain changes Trend and Persistence; REC pairs ABA Persistence. Outside N-mode, a selected HUD updates the rings to its bucket, while Trend stays fixed to the window. The Persistence ring shows the percentage of qualified observations above +3 ΔYears. This is a descriptive BioMIR model threshold, not a clinical event or a validated clinical risk threshold. Its heading explains the measure; it does not select bands. Trend describes direction, while Persistence describes frequency above that level. A declining series can remain elevated for much of a window. Missing observations are excluded; held inputs are not new measurements. The Adherence ring offers Completed Actions (%) by default and Reclaimed ΔYears (%). Completed Actions counts completed eligible actions over all eligible actions. Reclaimed ΔYears weights earned credit against eligible opportunity for the same day. Unresolved actions earn zero and stay in the denominator; unavailable support is not zero. These are records of reported follow-through and modeled credit, not independently verified behavior or measured age reversal. The REC detail Adherence table uses Band / ΔYears / N: ΔYears gives each band interval and N counts qualified opportunity days. The Summary separately reports action completion and reclaimable/reclaimed opportunity. REC does not define biological-band Persistence. Its companion Trend remains signed daily earned ΔYears credit across the visible window; scrubbing updates the period rings without refitting that window Trend. KDM/PhenoAge companion Trend and Persistence use actual panels in the trailing five years. These are descriptive records, not validated clinical outcomes.

Alongside the rings, Trend fits a straight least-squares line through the inspected domain’s displayed values within the main chart’s visible window. Values receive equal weight; the line spans the first to last supported dates. The fitted levels remain on the ΔYears scale, not a series of annualized rates. The β₁ disclosure separately expresses the fitted slope in ΔYears per calendar year, using 365.2425 days per year. Aggregate detail plots use their selected-period observations, with the clinical five-year exception below; their slope disclosure uses the same annual units. An absolute annualized slope below 0.0005 is displayed as neutral. This product rounding rule is neither a significance test nor a clinically validated change threshold. Scrubbing changes the period rings, while the regression remains fixed for that domain and window. Missing support stays unavailable, carried inputs do not become independent measurements, and neither a band share nor a fitted direction establishes disease risk, treatment effect or a biological aging rate. Read the definitions and interpretation guide.

Domain and biomarker longitudinal context

Domain Summary cards include the same trend and persistence presentation as biomarker detail, separated from summary values by a faint divider. Domain tables use stored domain ΔYears; percentages are not sums of biomarker persistence. Aggregate statistics retain their selected-period plot and tier counts except for the clinical five-year context. KDM and PhenoAge statistical details use trailing five years; their Day/N histories retain calendar 1Y, 3Y, 5Y and 10Y rolling windows. CMA retains 7D, 30D, 90D and 1Y rolling context. KDM and PhenoAge summaries count distinct persisted clinical-anchor dates, while daily Cardiometabolic context counts modeled days, including held measurements. These denominators are different and do not imply additional laboratory observations. Missing history stays unavailable; a ten-year window does not guarantee ten-year coverage.

Clinical history windows

KDM and PhenoAge Day/N histories use 1Y, 3Y, 5Y and 10Y lookback windows. Their statistical detail uses five years ending at the selected cutoff; clinical companion Trend/Persistence use five years ending at the visible main-window cutoff, capped at today. Contributors show distinct sampled calendar years as Yrs; panel-based regression and band counts retain their actual dated observations. Adherence remains selected-bucket REC. See the clinical five-year interpretation guide for Mean, Median, K and denominator limits. These are descriptive product choices, not recommended testing intervals or a clinically validated cadence. CMA, including FBG, BMI and blood pressure, retains 7D, 30D, 90D and 1Y rolling windows; nonclinical aggregate STAT views retain selected-period inspection. Each clinical rolling window displays its qualified observation count under History (N), without date ranges. A single observation cannot support a fitted trend; missing years are not filled.

KDM and PhenoAge biomarker persistence counts distinct qualified recorded dates, never repeated held results. KDM and PhenoAge domain histories count distinct persisted anchor dates. Cardiometabolic domain persistence remains a modeled-day summary using its daily rolling table.

Shorter inspection remains important: NIDDK recommends A1C testing at least twice yearly for people with diabetes; the 2026 ACC/AHA dyslipidemia guideline describes lipid monitoring 4–12 weeks after treatment initiation or adjustment and every 6–12 months thereafter. These sources support retaining recent-period inspection; they do not validate BioMIR’s four lookback windows or modeled persistence thresholds.

Four evidence classes

Read each claim at the appropriate level.

01

Published method

Klemera–Doubal Method (KDM) and Levine PhenoAge (clinical) are established published methods. Their original literature defines the mathematical and outcome-linked frameworks.

02

Supporting evidence

Meta-analyses, pooled cohorts, trials, and measurement-validation studies provide context for individual biomarkers and constructs. Association does not establish an individual causal effect.

03

BioMIR specification

Input selection, reference handling, directionality, standardization, weighting, aggregation, Δ-years translation, classifications, persistence, and missing-data behavior are implementation choices.

04

Validation status

The exact daily composite, coefficient set, classifications, and changes in Δ-years require separate evaluation for calibration, reliability, subgroup performance, reproducibility, and clinically meaningful outcomes.

Explore with repeatable Demo Data

Try synthetic scenarios to learn how changes in behavioral, functional, and cardiometabolic data appear across Today, Dashboard, and Trends—without using personal health records. Repeatable examples let you revisit the same history and compare what each view reveals.

The app also provides validation controls for checking consistent software behavior. These are software checks, not evidence of clinical validity or a measured change in aging. Detailed control procedures are available in the app’s Resource guides.

How we design the simulations

The simulations are designed around behavioral, biological, and clinical realism. Each starts with a plausible change in daily behavior, models how relevant functions and biomarkers may respond over time, and shows how those changes can appear in later clinical measurements. The design draws on published research for the pathways. The size and timing of the changes are simulation assumptions.

For example, Alcohol + Sleep models increased alcohol exposure, reduced total sleep, and a delayed glucose response. Sedentary models reduced daily movement, then lower aerobic fitness and a gradual rise in BMI. Stable Baseline shows ordinary variation without a programmed directional change. See the published studies on alcohol and sleep, sleep and insulin sensitivity, and step reduction and metabolic response.

The repeatable simulations include Stable Baseline, Cardio Risk, Alcohol + Sleep, Sedentary, Fitness Gain, Carb Transition, and Cardio Reset. Alcohol + Sleep links alcohol exposure to reduced total sleep and delayed glycemic change. Sedentary and Fitness Gain contrast activity trajectories, Carb Transition isolates dietary carbohydrate and glycemia, and Cardio Risk and Cardio Reset explore adverse and favorable cardiometabolic changes.

Cardio Risk and Fitness Gain are accelerated illustrations with terminal Adaptive BioAge targets above +8 and below −10 ΔYears, respectively. In intermediate scenarios, later responses can overlap as the modeled physiology changes. Directional cases require KDM and PhenoAge to move as expected across six dated specimen pairs. A 25% paired-clock difference is an advisory target when both offsets exceed one year in magnitude.

All seven scenarios passed software checks across three profiles. The result supports software verification. Clinical validity and prognostic performance require separate evidence.

Read the proposed UK Biobank external validation study for the next clinical-clock research step.

Personalization architecture

Keep intelligence downstream of the science.

BioMIR separates the reproducible model that generates the scientific result from downstream adaptive and generative systems that may personalize longitudinal interpretation or provide natural-language explanation and interaction.

Current · deterministic core

Prespecified calculations, versioned parameters, contributor attribution, provenance, and explicit data-support rules preserve reproducibility and make model changes identifiable rather than implicit.

Roadmap · private on-device learning

A downstream personalization layer could learn an individual’s baseline, intra-individual variability, temporal response, and persistent response patterns from longitudinal data while leaving the authoritative BioMIR equations unchanged.

Research direction · federated learning

Privacy-preserving federated approaches could eventually improve downstream population models across consenting users without centralizing raw longitudinal health records. Any such implementation would require separate privacy, security, scientific, and performance validation before release.

Readiness · constrained AI explanation

BioMIR’s staged Apple Intelligence architecture keeps scientific authority in deterministic engines, then exposes bounded structured results through a semantic/API layer and canonical BioMIR knowledge. Future Foundation Models, App Intents, voice, or Siri-facing components may explain, synthesize, narrate, and orchestrate those outputs, but they may not independently recalculate or silently alter the scientific result.

Implementation boundary: on-device adaptive learning, federated learning, and conversational AI are staged roadmap, research, or readiness directions—not capabilities of the current released app. Learned and generative layers are intended to operate on or interpret structured deterministic outputs rather than silently rewrite the scientific core. AI remains additive: loss or unavailability of a language model must not impair normal BioMIR computation or visualization.

Trace the evidence

Read the specification.
Read the sources.

Use Science for interpretation and Resources for curated literature and acquisition context.

Science & methodology Reference library