Elite Athletes and New Mothers Share a Physiology

Summary questions

Can wearable data really tell what I did during the day?

Yes, and more specifically than most people expect. Terra's internal health language model can identify from wearable signals alone that a user was drinking and partying, menstruating, or even received a massage that day. The model translates health-space patterns into natural language descriptions without being explicitly told what happened.

Can wearable data really tell what I did during the day?

Yes — surprisingly well. Terra's internal health language model was able to infer events like drinking, menstruation, and even getting a massage purely from wearable signals. The model translates health data into a representation space where similar physiological patterns cluster together, allowing it to map biometric shifts back to real-life behaviors and states.

Why would a health model confuse an elite male athlete with a postpartum woman?

Because their physiological signatures are genuinely similar. When measuring cosine similarity to the postpartum cluster centroid, male athletes scored 0.20 — higher than typical women (0.16) and the general population (0.12), and closest to actual new mothers (0.36). The model wasn't malfunctioning; it was detecting a real overlap in recovery physiology.

Why would a health model confuse an athletic man with a postpartum woman?

Because their physiological signatures are genuinely similar. When Terra measured cosine similarity to the postpartum cluster centroid, male athletes scored 0.20 — higher than typical women (0.16) and the general population (0.12), and second only to actual new mothers (0.36). The model wasn't malfunctioning; it was detecting a real overlap in health data.

What do new mothers and elite athletes actually have in common physiologically?

Both show lower resting heart rates, higher heart rate variability (HRV), and elevated calorie burn. Combined with fragmented sleep, these markers create a health signature the model reads as either heavy training load or postpartum recovery. Metabolically, a recovering new mother's body is doing work comparable to an athlete's.

How is a new mother's physiology similar to an elite athlete's?

Postpartum women show lower resting heart rates, higher heart rate variability (HRV), and higher calorie burn — the same markers that define elite endurance athletes. Combined with fragmented sleep from caregiving, the biometric fingerprint looks strikingly like a young man pushing his body hard in training. The recovery demands of new motherhood produce an athletic-grade physiological load.

How does a health language model figure out what's happening in someone's life?

It places people into a representation space where individuals with similar physiological patterns cluster together. Terra's model then maps regions of that space back to real-world experiences — like caregiving, menstruation, or intense training. Similarity to a known cluster centroid is what lets the model generate specific natural-language descriptions of a user's day.

How close is a new mother's health signature to the postpartum cluster compared to everyone else?

Real new mothers scored 0.36 in cosine similarity to the postpartum centroid, roughly 3x the general population's 0.12 and more than 2x typical women at 0.16. Male athletes landed at 0.20 — closer to postpartum than any other non-postpartum group tested. The gap between new mothers and everyone else is large, but athletes are the closest neighbors.

Does the model use gender or sex as an input?

No. Terra's datasets are gender and sex-neutral, and the expectation is that the model learns sex-linked health patterns implicitly from the physiological data itself. This mostly works — which is exactly why the athletic-male-to-postpartum-woman confusion was informative rather than just an error to patch.

Does this mean postpartum recovery is as demanding as elite training?

The data suggests yes, at least physiologically. If a model trained to distinguish subtle health states places postpartum women and male athletes in nearly the same region of representation space, the underlying cardiovascular and metabolic load is comparable. New mothers are effectively undergoing an athletic recovery process — just without the training plan or the recognition.

Why am I getting a 'postpartum' or 'caregiving' reading when that doesn't apply to me?

If you're training hard and sleeping poorly, your wearable signature — low resting HR, high HRV, elevated calorie burn, fragmented sleep — genuinely overlaps with postpartum physiology. The model described one young athletic man's day as "fragmented, broken sleep consistent with caring for a baby (with a pregnancy/postpartum signature)." It's not a random hallucination; it's two distinct life states producing near-identical biometrics.

Why do health language models make mistakes like this?

Because they operate in a representation space where physiologically similar people cluster together, regardless of demographic labels. Terra's datasets are gender- and sex-neutral by design, so when two very different life situations produce nearly identical biometric patterns — low resting HR, high HRV, fragmented sleep, elevated calorie burn — the model groups them. These edge cases aren't failures so much as revelations about physiology.

What does this failure case reveal about wearable-based health inference?

That physiology, not labels, is what wearables actually see. Two very different life situations — elite training and postpartum recovery — can occupy nearly the same region in health-representation space. Rigorous failure testing on models like this doesn't just improve accuracy; it surfaces genuine biological insights that wouldn't be obvious from either population studied in isolation.

What can wearable data reveal that I might not expect?

Far more than step counts. Terra's model surfaces states like partying, menstruation, massages, and postpartum recovery directly from wearable signals — and it does so by learning implicit patterns rather than being told what to look for. The interesting discoveries often come from the model's mistakes, which expose real physiological overlaps between seemingly unrelated groups.

How similar are postpartum women's biometrics to the general population?

Not very. New mothers sit at 0.36 cosine similarity to the postpartum centroid, while the general population sits at just 0.12 — a threefold gap. Postpartum recovery is a physiologically distinct state that clearly separates from baseline, which is why the model can identify it and why its overlap with elite male athletes at 0.20 is so striking.