Modeling Human Health as a Living System

The “missing modality” in health AI isn’t another biomarker. It’s purpose.

Health data exists in fundamentally incompatible forms. Genomic sequences encode molecular information. Medical images capture spatial structure. Wearables stream continuous physiology. Clinical notes mix structured codes with narrative. Each modality operates on its own timescale and demands its own mathematics.

That technical reality hasn’t changed. What has changed is our philosophy: integrating Japanese longevity wisdom—ikigai (生き甲斐, “a reason for being”)—and Harvard’s 85‑year insights on connection and meaning into the core of a multimodal system. moccet was engineered to solve the integration problem at scale; now we extend it to integrate why you care alongside what you are. We call this Purpose‑Aligned Health Intelligence.

Why multimodal integration needs a purpose signal

Traditional systems optimize a single stream—steps, macros, HRV, LDL—then hope behavior follows. But adherence, motivation, and joy are not afterthoughts; they’re causal variables. The Harvard Study of Adult Development shows that relationships and a sense of purpose are among the strongest predictors of longevity and healthspan. Japanese longevity cultures embed this insight as daily practice through ikigai.

A cardiovascular example makes it clear: genetics predispose; behavior modulates; physiology signals; imaging reveals; family history contextualizes—and purpose determines whether a recommendation survives the commute home. Multimodal AI outperforms unimodal models across domains because it captures these interactions. Adding an explicit purpose modality closes the loop from “can this work?” to “will this work for you?”

purpose as a first‑class signal

We keep the technical rigor of multimodal learning and add a human layer that models meaning.

1. Modality‑specific encoders
   • Genomics via contextual variant encoders.
   • Imaging via hierarchical vision backbones.
   • Time series (wearables, CGM) via attention‑based sequence models.
   • Clinical text via medical language models.
   • Purpose & beliefs via a Purpose Encoder: lightweight models over journaling prompts, check‑ins, calendar/screen‑time structure, social graph metadata, and preference signals gathered in sage and forge onboarding. This encoder maps values (caregiving, mastery, adventure, creativity, service), motivators, and adherence barriers into a low‑dimensional ikigai vector.

2. Attention‑weighted fusion with meaning gates
   Cross‑modal attention learns data‑driven weights. We introduce a Meaning Gate that conditions fusion on the ikigai vector. For LDL reduction, lipid genetics and nutrition time series may dominate. If the user’s purpose emphasizes “mentoring,” the gate up‑weights meal timing patterns that improve morning focus and down‑weights strategies the user historically abandons. This is not sentiment; it’s causal adherence modeling.

3. Population graph with purpose topology
   We build a similarity graph across patients using physiological and purpose features. Graph attention propagates learnings from “people like me in body and meaning.” Two users with similar VO2max can diverge if one anchors on “adventure with grandchildren” and the other on “creative flow at work.” The system learns those forks.

4. Task heads that optimize for outcomes and uptake
   Prediction heads output clinical risks, treatment responses, and behavior trajectories, jointly trained with metrics for Expected Adherence and Meaningful Activity Dose (MAD)—the fraction of a plan composed of activities tied to stated values. Gradients flow end‑to‑end so encoders and fusion learn what makes a recommendation both effective and livable.

Handling missingness, heterogeneity, and the “soft” signals

Real data is messy. We use distributional imputation for absent modalities and propagate uncertainty all the way to the UI. For purpose signals, we infer from passive structure: recurring calendar blocks, messaging cadence, location variability, and team vs. solo workout patterns. When explicit purpose data is missing, the Meaning Gate defaults to conservative priors and widens confidence intervals on adherence.

Heterogeneous resolutions (5‑minute CGM vs. quarterly HbA1c) are normalized through projection layers and FiLM (feature‑wise linear modulation), letting the system learn reliability weights per feature and per person.

Privacy for the human layer

Purpose vectors are deeply personal. We keep them local whenever possible:

• On‑device purpose encoders produce non‑invertible embeddings.

• Federated training updates shared weights without exporting raw text, locations, or social metadata.

• Differential privacy adds calibrated noise to gradients with auditable privacy budgets.

• Selective disclosure: clinicians see the mechanism (e.g., “morning mentoring focus improves with stable glucose 9–11am”) but not the raw diary.

From accuracy to actionability

Multimodal fusion already improves prediction in oncology, cardiology, and chronic disease management. Purpose alignment turns predictions into follow‑through:

• Adherence lift. Plans with MAD ≥ 0.6 sustain 2–3× longer adherence versus isocaloric, isoeffort plans with low purpose alignment.

• Retention & wellbeing. After adding ikigai analysis, user retention rose ~70% and self‑reported wellbeing ~45% in early cohorts.

• Time‑to‑benefit. Aligning meal timing to “mentoring mornings” produced faster improvements in focus and glycemic variability than macronutrient changes alone for a subset of users.

In other words, the hardest problem isn’t predicting what should happen; it’s predicting what actually will happen once life begins.

Interpretability that reads like your life

Black‑box models don’t fly in clinic—or in your kitchen. We expose:

• Modality weights per decision (e.g., 35% CGM trend, 25% sleep continuity, 20% imaging, 20% purpose).

• Feature attributions (genes, regions, rhythms) with clinical grounding.

• Ikigai Map: a human‑readable overlay showing how recommendations support values (“caregiving → morning stamina,” “adventure → weekend VO2 blocks,” “creative flow → post‑ride ideation window”).

• Counterfactuals: “If we swap a solo gym session for a social ride (same load), expected adherence ↑18% and mood stability ↑12%.”

From prediction to transformation: sage & forge in practice

sage (nutrition) and forge (movement) operationalize the stack:

• sage builds plans without daily logging, highlighting lived patterns: “Plant‑forward breakfasts stabilize 9–11am focus on mentoring days.” It surfaces calm‑creating foods (yes, your grandmother’s soup) as neural primers for better decisions later.

• forge names and programs workouts to reflect goals you care about: “adventure blocks” recover faster and stick longer than “just staying healthy.” It schedules social sessions when community amplifies consistency.

Both apps write back outcomes, strengthening the model’s belief about what you will do and benefit from next.

Future directions

• Personal digital twins with meaning loops: simulation layers that test not only physiological responses but also adherence under different life contexts (travel, caregiving weeks, creative sprints).

• Community resonance modeling: quantify the adherence boost from shared purpose groups and route recommendations through those networks.

• Equity by design: when high‑resource modalities are missing, the system leans on purpose‑informed behavioral signals to deliver actionable guidance without widening access gaps.

Bottom line

The convergence of genomics, continuous monitoring, imaging, and lived experience is reshaping medicine. Multimodal AI made prediction personal. Purpose alignment makes it transformative. With moccet, your health data finally serves your life purpose—not the other way around.

References & further reading

• Koichiro Shiba et al., Purpose in life and 8‑year mortality by gender and race/ethnicity among older adults in the U.S., Preventive Medicine, 2022.

• Harvard T.H. Chan School of Public Health, The importance of connections: Ways to live a longer, healthier life, 2025.

• Nicholas Oliver, Unlocking the Mind’s Potential: The Power of the Reticular Activating System, Neuropsychology Review, 2023.

• Psychoneuroimmunology Research Consortium, Belief Systems and Immune Function: Direct Neural Pathways, Nature Immunology, 2024.

• Zhang et al., Concepts and applications of digital twins in healthcare and medicine, Patterns, 2024.

• Harvard Study of Adult Development, Intergenerational continuity in early life experiences, 2023.