Turning Health Data Into
Actionable Intelligence
A classical ML pipeline that discovers multi-system health patterns from longitudinal patient records — designed to be validated on 500,000+ patient records, exportable as non-generative models.
5 Core Methods
Non-generative, classical ML — fully exportable
500K–800K
Patient cohort from Mayo's 2.5–3M records
10–20+
Overlapping phenotypes discovered from data

Every patient is a complex system — metabolic, inflammatory, sleep, renal, hepatic, cardiovascular, and nutritional systems interact continuously. Today's clinical tools analyze these in silos. Healthena's phenotyping engine connects them: we encode raw labs, vitals, and medications into discrete clinical states, align them to a common weekly timeline with confidence-weighted carry-forward, then run two parallel analyses — fuzzy clustering discovers overlapping health patterns, and lagged cross-correlation maps directional system-to-system influence. The outputs feed into calibrated risk predictors (XGBoost + Cox PH) targeting four longevity-relevant outcomes: cluster transition, new chronic diagnosis onset, cardiometabolic deterioration, and PhenoAge acceleration. Every component is non-generative and exportable.

The Challenge

The Problem

Healthcare today analyzes the human body in silos. A patient's metabolism, sleep, inflammation, kidney function, and biological aging are managed by different specialists, in different visits, looking at different metrics — but chronic disease doesn't respect silos. Sleep dysregulation affects glucose. Systemic inflammation affects recovery. Lipid dysfunction affects cognition. By the time a disease is diagnosed, multi-system damage has already accumulated for years.

Longevity and preventative-care clinics need to see the whole picture — but no clinical tool today integrates labs, vitals, medications, and trajectory into one cohesive view of how a patient is aging across all systems simultaneously. Existing tools either focus on single-disease risk (cardiovascular, diabetes) or produce dashboards of disconnected numbers without revealing the underlying phenotype.

Our Approach

Our Solution

The Multi-System Phenotyping Engine integrates labs, vitals, medications, and diagnoses across all physiological systems into discrete clinical states, aligns them on a common weekly timeline with confidence-weighted carry-forward, and uses fuzzy clustering to discover overlapping multi-system phenotypes directly from data — not from hand-named hypotheses.

The engine maps directional influence between systems (which systems predict which other systems' future states), assigns each patient a continuous membership vector across all discovered phenotypes, and predicts four longevity-relevant outcomes: phenotype transition from healthy to disease, new chronic diagnosis onset, cardiometabolic deterioration, and biological aging acceleration.

Every component is non-generative classical machine learning — fully exportable, auditable, and deployable in clinical settings.

Healthena Multi-System Phenotyping Engine

A non-generative classical ML pipeline for discovering and validating multi-system health phenotypes from longitudinal EHR data. Every output artifact is a learned parameter, threshold, or graph weight—never free-text generated by an LLM.

Pipeline Architecture

INPUT PROCESSING PATTERN DISCOVERY OUTPUTS Raw Data Labs • Vitals Medications • Dx Demographics State Encoding Discretize to −1 • 0 • +1 per metric Time Alignment Confidence-weighted carry-forward weekly grid alignment Fuzzy Clustering Health pattern discovery (FCM) 10–20+ phenotypes Lagged Cross-Corr Directed system influence graph p < 0.01 edges only Outcome Prediction XGBoost + Cox PH 1. Cluster transition (Healthy → disease) 2. New chronic dx onset 3. Cardiometabolic deterioration 4. PhenoAge accel. (≥2 yr / 12 mo) DATA FLOW Raw observations → discretized into clinical states → aligned to weekly grid with confidence decay → fed into two parallel pattern-discovery branches (clustering + system correlations) → outputs converge into calibrated outcome predictions.
Pipeline Summary

State Encoding converts raw clinical observations (labs, vitals, medications, diagnoses) into a discrete ternary representation ($-1$, $0$, $+1$) anchored to evidence-based clinical thresholds, enabling cross-metric comparison without scale harmonization.

Time Alignment projects heterogeneously-sampled observations onto a regular weekly grid using exponential confidence decay—the observed value is carried forward unchanged, but the system's confidence in that value decays with a modality-specific half-life, formally separating "what we know" from "how sure we are."

Parallel Analytic Branches extract two complementary structures: (a) Fuzzy C-Means clustering discovers overlapping multi-system health patterns; (b) Lagged cross-correlation maps directional influence between physiological systems. These converge into calibrated risk predictions (XGBoost + Cox PH) for four longevity-relevant outcomes.

1 State Encoding

The first transformation converts heterogeneous clinical observations into a discrete, comparable representation. Each metric is mapped to one of three states using its clinically established reference range as the decision boundary.

Threshold-Based Encoding

For metric $m$ with observed value $x$ at time $t$, and clinical thresholds lower bound $L_m$ and upper bound $U_m$:

$$s_t^m = \begin{cases} -1 & \text{if } x < L_m \quad \text{(below normal)} \\ \phantom{-}0 & \text{if } L_m \le x \le U_m \quad \text{(normal)} \\ +1 & \text{if } x > U_m \quad \text{(above normal)} \end{cases}$$

For metrics without established clinical thresholds (e.g., HRV, sleep regularity), we fall back to cohort-percentile bucketing:

$$s_t^m = \begin{cases} -1 & \text{if } x \le P_{25}(\text{cohort}) \\ \phantom{-}0 & \text{if } P_{25} < x < P_{75} \\ +1 & \text{if } x \ge P_{75} \end{cases}$$
State Encoding Is Not a Bottleneck

The {−1, 0, +1} encoding looks like it loses information, but it doesn't. State encoding is purely a normalization tool for cross-modality comparability in the clustering step. The patient summary vector that flows into clustering and prediction retains all the continuous information:

  • Carried numeric value with confidence weight (from Section 2)
  • 1-month, 3-month, 12-month trajectory deltas (continuous)
  • Volatility / standard deviation (continuous)
  • Fraction of weeks in each state (continuous)
  • Longest abnormal streak (continuous)

The discrete state is one feature among many. Downstream models see the rich continuous signal.

Encoding Demonstration

Metric Raw Value Lower Threshold ($L_m$) Upper Threshold ($U_m$) State ($s_t^m$) Interpretation
LDL Cholesterol150 mg/dL—130 mg/dL+1Above optimal
HbA1c6.3%—5.7%+1Pre-diabetic range
hsCRP1.5 mg/L—3.0 mg/L0Normal range
eGFR85 mL/min60 mL/min—0Normal renal function
Vitamin D16 ng/mL30 ng/mL—-1Deficient
Albumin4.2 g/dL3.5 g/dL5.5 g/dL0Normal range
Worked Example — Patient "Sarah Chen"

Sarah Chen, age 45, female. Eight key metrics encoded from her most recent lab panel:

MetricRaw ValueThreshold(s)State
LDL150 mg/dLU = 130+1
ApoB120 mg/dLU = 100+1
HbA1c6.3%U = 5.7+1
Fasting Glucose122 mg/dLU = 100+1
hsCRP1.5 mg/LU = 3.00
eGFR85 mL/minL = 600
Vitamin D16 ng/mLL = 30-1
Albumin4.2 g/dLL=3.5, U=5.50

State vector: $\mathbf{s}_{\text{Sarah}} = [+1, +1, +1, +1, 0, 0, -1, 0]$

Four of eight metrics are in the abnormal range, all concentrated in the metabolic domain (LDL, ApoB, A1c, glucose). This pattern will drive high membership in the Metabolic-Driven cluster in Section 3. The Vitamin D deficiency ($-1$) will contribute to the Nutritional-Deficient dimension. Notably, inflammatory (hsCRP) and renal (eGFR) markers are normal, which constrains the differential.

Why ternary and not continuous? The ternary encoding is intentional. Cross-correlation on system-level states and fuzzy clustering on summary features both benefit from this discretization. By quantizing at clinical thresholds rather than arbitrary percentiles, we preserve clinical interpretability: a state of $+1$ for LDL means the same thing for every patient. Continuous z-scores are retained for the outcome prediction module, which operates on the unquantized data where appropriate.

2 Time Alignment with Confidence Weighting

The VALUE stays unchanged. What decays is CONFIDENCE. A patient's last measured LDL of 150 mg/dL does not drift toward some prior over time—it remains 150. What changes is the system's certainty that 150 still reflects reality. This distinction is critical: clinicians reason about stale labs the same way ("her LDL was 150, but that was 8 months ago").

Carried Value

Let $t' < t$ be the most recent measurement time for metric $m$ before grid point $t$. The carried value is:

$$x_{\text{carried}}^m(t) = x_{t'}$$

The value is unchanged—no interpolation, no extrapolation.

Confidence Decay

$$c_m(t, t') = \exp\!\left(-\frac{t - t'}{\tau_m}\right)$$

where $\tau_m$ is the modality-specific half-life time constant.

Optional Shrinkage to Prior

$$x_{\text{aligned}}^m(t) = c \cdot x_{t'} + (1 - c) \cdot \mu_m$$

Modality Half-Lives

Modality Decay constant (days) Half-life (days) Rationale
Wearable vitals (HR, HRV, SpO2)74.9High-frequency signal; physiology changes within days
Sleep metrics149.7Weekly patterns are stable; seasonal drift over ~2 weeks
Blood pressure3020.8Responds to meds/lifestyle in 2–4 weeks
Metabolic labs (glucose, A1c, lipids)180124.7A1c reflects ~90-day average; lipids are stable absent intervention
Renal markers (eGFR, creatinine)180124.7Chronic kidney trajectory is slow-moving
Inflammatory markers (hsCRP, ESR)9062.4Reactive in acute illness but baseline is relatively stable
CBC / hematology12083.2Moderate turnover; anemia develops over months
Thyroid (TSH, T4)180124.7Slow-moving endocrine axis
Nutritional (Vit D, B12, folate)180124.7Fat-soluble vitamins have long half-lives in tissue
Diagnoses / medication changes∞ (never decays)∞Binary facts — a diagnosis is forever until explicitly resolved

Time Alignment Visualization

Worked Example — Sarah's LDL Timeline

At month 10 (between measurements at month 6 and month 14), the system needs an LDL value. The most recent measurement is 148 at month 6.

Carried value: $x_{\text{carried}}^{\text{LDL}}(10) = 148$ (unchanged)

$$c_{\text{LDL}}(10, 6) = \exp\!\left(-\frac{120}{180}\right) = \exp(-0.667) \approx 0.513$$

The system carries forward 148 mg/dL with 51.3% confidence. Once a new measurement arrives at month 14 (155), confidence resets to 1.0.

3 Patient Feature Summarization + Fuzzy C-Means Clustering

Critical Distinction

FCM does NOT receive the raw weekly state vectors. Weekly states are variable-length time series — FCM requires a fixed-length vector per patient. We first summarize each patient's weekly history into a fixed-length feature vector (~80 dimensions), then cluster those summaries.

The raw weekly state vectors are used directly by Lagged Cross-Correlation (Section 4). FCM needs a snapshot of where the patient is, how they've trended, and how volatile they are.

Step 3a — Summarize Weekly States into Patient Feature Vectors

FeatureDescriptionExample (Sarah's LDL)
current_state_mMost recent week's state+1 (above)
mean_state_mMean state over last 12 weeks+0.9
trend_mSlope of states (improving/stable/worsening)0.0 (stable high)
volatility_mStd dev of states0.1
frac_above_mFraction of weeks in above-normal state0.95
frac_below_mFraction of weeks in below-normal state0.0
longest_streak_mLongest consecutive weeks in abnormal state12

Step 3b — Fuzzy C-Means on Summary Vectors

Unlike hard clustering (K-means), Fuzzy C-Means (FCM) assigns each patient a continuous membership vector across all clusters. A patient can be 72% Metabolic-Driven and 15% Inflammation-Linked simultaneously. This matches clinical reality: multi-morbidity is the norm, not the exception.

Objective Function

$$J(U, C) = \sum_{i=1}^{N}\sum_{k=1}^{K} (u_{ik})^p \cdot \|x_i - c_k\|^2$$

Update Rule 1: Centroids

$$c_k = \frac{\sum_{i=1}^{N} (u_{ik})^p \cdot x_i}{\sum_{i=1}^{N} (u_{ik})^p}$$

Update Rule 2: Memberships

$$u_{ik} = \frac{1}{\displaystyle\sum_{j=1}^{K}\left(\frac{\|x_i - c_k\|}{\|x_i - c_j\|}\right)^{2/(p-1)}}$$

Optimal $K$ Selection

$$K^* = \arg\max_K \Big[\text{Silhouette}(K) - \alpha \cdot \text{XieBeni}(K)\Big]$$

Cluster Feature Space

Discovered Health Pattern Taxonomy ($K^* = 12$)

❤️
1. Metabolic-Driven
LDL, ApoB, glucose, HbA1c, BMI
18% activation
💤
2. Sleep-Recovery Constrained
HRV, RHR, sleep duration, sleep regularity
14% activation
🔥
3. Inflammation-Linked
hsCRP, NLR, ferritin, ESR
12% activation
🪷
4. Renal-Stressed
eGFR, ACR, creatinine, BP
9% activation
🫁
5. Hepatic-Strained
ALT, AST, GGT, platelets
7% activation
🩸
6. Vascular-Stiffening
Pulse pressure, lipids, age, PWV
8% activation
⚖️
7. Endocrine-Imbalanced
TSH, glucose, thyroid meds, cortisol
6% activation
🍊
8. Nutritional-Deficient
Vit D, B12, folate, iron, albumin
10% activation
🩸
9. Anemia / Hematologic
Hgb, MCV, iron, B12, ferritin
5% activation
🧠
10. Cognitive-Risk
Vit D, B12, homocysteine, glucose variability
4% activation
🛡️
11. Autoimmune-Active
CRP, CBC, Dx codes, immunosuppressants
3% activation
✅
12. Stable-Healthy
Reference cluster — all metrics normal range
22% activation

Worked Example — Patient Membership Vectors

4 System Correlations (Lagged Cross-Correlation Graph)

Lagged cross-correlation measures the linear association between the current state of system $X$ and the future state of system $Y$, computed at actual measurement timepoints. Unlike standard correlation, the lag structure makes this measure directed — $r_{XY}(\text{lag}) \neq r_{YX}(\text{lag})$ — so we can determine which system predicts which.

Lagged Cross-Correlation at Measurement Points

$$r_{XY}(\text{lag}) = \text{Corr}\!\big(X(t),\; Y(t + \text{lag})\big)$$

Directed Edge Retention

A directed edge $X \to Y$ is retained if: (1) $r_{XY}(\text{lag}=1)$ is significant at $p < 0.01$; (2) $|r_{XY}| > |r_{YX}|$; (3) Edge weight $= |r_{XY}(\text{lag}=1)|$.

System Correlation Graph

Sleep HRV, duration Recovery RHR, HRV trend Metabolic Glucose, A1c Inflammatory hsCRP, NLR Lipids LDL, ApoB, TG Renal eGFR, ACR 0.42 0.31 0.28 0.22 0.19 0.24 0.17 0.14 Arrow thickness = |r(lag=1)| | All edges significant at p < 0.01 (permutation test, n=1000) | Directionality: |r_XY| > |r_YX|
Clinical significance: The strongest edge in our graph is Sleep → Recovery ($r = 0.42$), reflecting the well-established physiological dependence of autonomic recovery on sleep quality. The Recovery → Metabolic edge (0.31) captures HPA-axis-mediated metabolic effects of poor recovery. These directed edges become explanatory context for the AI companion: "Your sleep disruption last week is likely contributing to this week's elevated glucose."

5 Outcome Prediction (XGBoost + Cox PH)

Feature Vector

$$\varphi_i = \Big[\underbrace{s_t^1, \ldots, s_t^M}_{\text{current states}},\; \underbrace{\Delta_t^{1m}, \Delta_t^{3m}, \Delta_t^{12m}}_{\text{trajectory deltas}},\; \underbrace{u_{i1}, \ldots, u_{iK}}_{\text{cluster memberships}},\; \underbrace{\text{age, sex, BMI, \ldots}}_{\text{demographics}},\; \underbrace{dx_1, \ldots, dx_D}_{\text{comorbidities}}\Big]$$

XGBoost Classifier

$$\hat{y}_i = f_{\text{XGB}}(\varphi_i; \hat{\theta}) \qquad \hat{p}_i = \text{isotonic}(\hat{y}_i)$$

Cox Proportional Hazards

$$\lambda(t \mid \varphi) = \lambda_0(t) \cdot \exp\!\left(\beta^T \varphi\right)$$

Model Calibration

Note on the numbers in these charts: The 0.81 AUC and the curve shapes shown below are illustrative targets, not actual results — we don't have Mayo data access yet. The charts show what these plots will look like when we have real results.

Calibration Plot (Cardiometabolic Deterioration Composite)

ROC Curve — Cardiometabolic Deterioration (AUC = 0.81)

Feature Importance (SHAP Values — Sarah Chen)

6 Calibration & Validation Metrics

Discrimination & Calibration Targets
OutcomeTarget AUCCalibration (Brier)Notes
Cluster transition (Healthy → disease)≥ 0.78≤ 0.15Requires sufficient Stable-Healthy patients with 12-month follow-up
New chronic diagnosis onset≥ 0.75≤ 0.12Lower prevalence event; Cox PH primary model
Cardiometabolic deterioration composite≥ 0.80≤ 0.18Composite outcome boosts sample size; XGBoost primary
PhenoAge acceleration (≥2 yr / 12 mo)≥ 0.76≤ 0.14Requires all 9 Levine labs at both timepoints
Validation Protocol
  1. Temporal split: Train on data before cutoff date, test on data after. No future leakage.
  2. 5-fold cross-validation within training set for hyperparameter tuning.
  3. Isotonic calibration on a held-out calibration fold (separate from test set).
  4. Subgroup fairness: AUC and calibration reported stratified by age band, sex, and race/ethnicity. Maximum AUC disparity across subgroups ≤ 0.05.
  5. Confidence intervals: 1,000-iteration bootstrap on test set for all metrics.

7 Five Patient Profiles

SC
Patient 1: Sarah Chen
Age 45 • Female • Asian • BMI 30.7 • No major comorbidities
Inputs

Demographics

Age
45
Sex
Female
Ethnicity
Asian
BMI
30.7

State Trajectory (5-Week Window)

MetricWk 1Wk 2Wk 3Wk 4Wk 5Trend
LDL (mg/dL)148 148 150 150 150 Stable-high
HbA1c (%)6.1 6.1 6.3 6.3 6.3 Rising
hsCRP (mg/L)1.2 1.4 1.5 1.6 1.5 Slightly rising
Vit D (ng/mL)16 16 16 16 16 Stable-low
Sleep (hrs)7.2 7.0 7.1 7.3 7.1 Stable-normal
Pipeline Outputs

Health Pattern Membership

Risk Scores (12-month horizon)

Cluster transition: 12% New chronic dx: 8% Cardiometabolic deterioration: 34% PhenoAge acceleration: 22%
Companion Recommendation
Top Concern: Lower cardiometabolic risk (glucose + LDL/ApoB trajectory) while protecting recovery baseline.
JM
Patient 2: James Mitchell
Age 52 • Male • White • BMI 26.2 • No major comorbidities
Inputs

State Trajectory (5-Week Window)

MetricWk 1Wk 2Wk 3Wk 4Wk 5Trend
Sleep Duration (hrs)7.4 7.3 7.1 7.2 6.8 Declining
HRV (ms)61 60 56 59 55 Declining
hsCRP (mg/L)2.2 2.4 2.6 2.7 2.8 Rising toward threshold
NLR2.5 2.7 2.9 3.0 3.1 Rising, crossed threshold
Pipeline Outputs

Health Pattern Membership

Risk Scores (12-month horizon)

Cluster transition: 18% New chronic dx: 15% Cardiometabolic deterioration: 19% PhenoAge acceleration: 28%
Companion Recommendation
Top Concern: Address sleep-recovery strain driving inflammatory elevation. HRV declining (61 → 55), NLR crossed threshold.
EN
Patient 5: Elena Nowak
Age 41 • Female • White • BMI 22.8 • No comorbidities
Inputs

State Trajectory (5-Week Window)

MetricWk 1Wk 2Wk 3Wk 4Wk 5Trend
LDL (mg/dL)105 104 105 106 105 Stable-normal
HbA1c (%)5.2 5.2 5.2 5.2 5.2 Stable-normal
HRV (ms)72 74 71 73 72 Stable-good
BP (mmHg)118/74 116/72 120/76 118/74 118/74 Optimal
Pipeline Outputs

Health Pattern Membership

Risk Scores (12-month horizon)

Cluster transition: 3% New chronic dx: 2% Cardiometabolic deterioration: 4% PhenoAge acceleration: 5%
Companion Recommendation
Status: All systems functioning well. Vitamin D at 32 is adequate but suboptimal — optimization opportunity only, not a clinical concern.

8 Data We Need From Mayo

Core Data (Standard EHR)

CategorySpecific DataWhy We Need It
Lab ResultsLipid panel (LDL, HDL, TG, ApoB, Lp(a)), HbA1c, fasting glucose, insulin, CMP, CBC w/diff, hsCRP, ferritin, ESR, vitamin D, B12, folate, iron, albumin, TSHState encoding inputs, phenotype features, outcome definitions, PhenoAge computation
Vitals (Flowsheet)HR, BP, weight, BMI, SpO2, walk time, grip strengthTime-alignment pipeline, system correlation inputs
DiagnosesICD-10 codes with onset datesCohort inclusion, comorbidity features, outcome event definitions
MedicationsDrug name, class, start/stop dates, doseTreatment-response features, phenotype context
DemographicsAge, sex, race/ethnicityDemographic stratification, fairness audits
EncountersVisit type, date, provider specialtyLongitudinal density, time-alignment anchor points
Social HistorySmoking, alcohol, exercise, sleep self-reportLifestyle features for phenotype discovery
Highest-Value Enrichment — TAME Geroscience Panel
  • Inflammaging: IL-6, CRP, sTNFR-II
  • GDF-15 (strongest single mortality predictor in St. Sauver 2023)
  • IGF-1 + fasting insulin (nutrient-signaling / longevity axis)
  • VO2 max (cardiorespiratory fitness) and grip strength
  • Cystatin-C and NT-proBNP (cardiovascular aging)
Standard clinical labs can discover metabolic, inflammatory, and organ-system phenotypes. But molecular senescence markers (GDF-15, SASP, P16) would enable discovery of longevity-specific phenotypes invisible to standard labs. We can deliver everything promised with standard EHR data. If aging-specific research data is available, we can discover phenotypes nobody else in the program is asking for.

9 End-to-End Pipeline

INPUTS (Mayo Data) WHAT WE DO (Processing Pipeline) OUTPUTS (What We Export) MAYO EHR DATA Labs Lipids, A1c, CBC, CMP, hsCRP, etc. Vitals BP, HR, weight, BMI, walk time Diagnoses ICD-10 codes with onset dates Medications Drug classes, start/stop dates Demographics Age, sex, race/ethnicity Encounters Visit type, dates, specialty Social / Family History Smoking, alcohol, exercise, family Hx [Optional] Molecular aging markers PROCESSING PIPELINE 1. State Encoding Discretize to {-1, 0, +1} 2. Time Alignment Weekly bins + confidence decay 3. Patient Summarization 80-dim feature vectors 4. Fuzzy C-Means Discover 10-20+ phenotype patterns 5. Lagged Cross-Correlation System correlation graph 6. XGBoost + Cox PH Outcome prediction models EXPORTED ARTIFACTS Cluster centroids + taxonomy JSON — phenotype definitions Patient membership vectors Per-patient cluster weights System correlation graph JSON — directed weighted edges XGBoost classifiers JSON — 4 outcome models Cox PH models Survival coefficients + baseline hazard Reference distributions Demographics + PhenoAge baselines SHAP feature importance Per-feature contribution weights Validation report AUC, calibration, fairness metrics Source code Python — full reproducible pipeline
What the companion does with these outputs post-program

The exported models become tools the personal health AI companion calls at runtime. Cluster centroids and reference distributions become RAG context. The system-correlation graph provides evidence-backed explanations. The companion itself is never fine-tuned on Mayo data — it gets smarter by operating over Mayo-validated tools and knowledge.

10 What Mayo Data Enables

Before vs. After Mayo
CapabilityBefore MayoAfter Mayo
Cluster discovery12 hypothesized phenotypesData-driven $K^*$ with validated centroids
System correlation graph6 edges based on literatureStatistically significant directed graph ($p < 0.01$, $n = 2.5$M)
Outcome predictionRule-based heuristicXGBoost + Cox PH with AUC > 0.80 and isotonic calibration
Confidence decayLiterature-based $\tau_m$Empirically calibrated half-lives per metric per population
Statistical powerCannot detect rare patterns0.5% support = 12,500 patients; detects small effects

Timeline

PhaseWeeksDeliverable
1. Cohort assembly & feature extraction1-6Cleaned, state-encoded, time-aligned dataset; ~80-dim summary vectors
2. Pattern discovery5-11Validated $K^*$, FCM centroids, lagged cross-correlation graph
3. Outcome modeling & calibration9-16Calibrated XGBoost + Cox PH for 4 outcome endpoints, fairness audit
4. Validation & export15-20Validation report; export package; integration into Healthena tooling

20 weeks of data access, from cohort assembly to an approved, validated export package.

11 AI Infrastructure Compliance

Generative imports: NONE

We do not bring in GPT-4o, Gemini, BERT, or any LLM. Our methodology is entirely non-generative by design. Our personal health AI companion stays on Healthena's own infrastructure and is never modified by Mayo data.

What We Export (Non-Generative Artifacts Only)
ExportMayo CategoryCompliance Basis
Trained XGBoost / LightGBM models (JSON)Traditional model exportTree-based, no embeddings, no gradients of source data
Trained Cox PH modelsTraditional model exportClassical statistical model, interpretable coefficients
FCM cluster centroidsTraditional model + aggregated statsCentroids are aggregates over many patients; no individual data
System-correlation graphAggregated statisticsPopulation-level graph, not per-patient
Validation report (AUC, Brier, calibration, fairness)Aggregated statisticsCohort-level metrics with CIs
Source codeSource codeReviewed for any embedded patient data — none
What we will NOT export
ItemReason
Any row-level or patient-level dataMayo policy prohibits
Embeddings, vectors, or gradients of source dataRestricted under Mayo's embedding export rule
Fine-tuned generative modelsNot allowed; also not applicable to our methodology
Any artifact with <10 patients per aggregated cellMayo policy minimum

12 References

Key References

St. Sauver et al. (2023). Biomarkers of cellular senescence and risk of death in humans. Aging Cell, 22, e14006.

Levine et al. (2018). An epigenetic biomarker of aging for lifespan and healthspan. Aging, 10(4), 573–591.

Bezdek, J. C. (1981). Pattern Recognition with Fuzzy Objective Function Algorithms. Plenum Press.

Chen & Guestrin (2016). XGBoost: A scalable tree boosting system. KDD 2016, 785–794.

Lundberg & Lee (2017). A unified approach to interpreting model predictions. NeurIPS, 30, 4765–4774.

Cox, D. R. (1972). Regression models and life-tables. J. Royal Statistical Society: Series B, 34(2), 187–202.

Collins et al. (2015). TRIPOD statement. BMJ, 350, g7594.

Several specific implementation choices (modality-specific confidence half-lives, K-selection trade-off parameter α, etc.) are starting values to be refined empirically against Mayo's cohort during the program. The references above ground the methodology; the parameters will be fit from data and reported in the final validation package.