Self-Supervised Behavioral Representations Across the Life Course: A Killifish Case Study

Self-supervised foundation models of aging are increasingly built from longitudinal data (biobanks, electronic health records, wearables) that is inherently incomplete: no individual is followed across a whole lifetime, and how much of each life is captured varies widely. This raises two linked questions: is it worth modeling an individual's whole life course rather than its current state, and can such a model be built from brief, fragmentary records? No human cohort can settle them, because none offers a complete life to compare against. We turn to the African turquoise killifish (Nothobranchius furzeri), tracked from youth to natural death in publicly released recordings, as a controlled testbed: its complete lifespans provide the full-life reference that human data lacks. On these data we build LifeMAE, a two-stage selfsupervised model: a day encoder that summarizes each day of behavior, then a life-course encoder over the trajectory of those daily summaries. We find that the day encoder alone is already strong: from a single day of behavior it predicts chronological age, separates long- from short-lived individuals (coarsely), and flags nearness to death. Adding the life-course encoder improves on none of the three; each is matched by trivially aggregating the day-level predictions (a smoother for age, an early-life average for lifespan). Near-term mortality seems the exception, where the whole-life model looks far better (AUROC 0.81 to 0.91), but the gain is not behavioral: it reflects where each day falls within the observation window (a cue supplied by the model's encoding of time), and a single-day model given that cue closes the gap at any observation length. For these traits, an individual's place in its life course is legible from a single day: the trajectory stage is unnecessary, and the record it needs is as short as one day, the finest grain our day-level setup resolves. For characterizing a cohort, this favors observing many individuals briefly over tracking a few for long. The result joins a growing body of work in which deep and foundation models, fairly benchmarked, fail to beat deliberately simple baselines. We add a concrete mechanism for the over-optimism: a model's encoding of time can leak the very quantity it predicts, which backwardlooking evaluation mistakes for learned biology, so only evaluation fixed to the moment of prediction is trustworthy.