The Science Behind Our Mood Tracking Algorithm

When we launched the first version of YouMindo's mood tracker in 2022, it worked like almost every other mood tracker on the market. You tapped a number from 1 to 10. We stored it. We showed you a line graph. That was it.

Two years of outcome data later, we know that a single daily number is a remarkably poor predictor of anything clinically meaningful. Here's what we did about it — and why the answer was more complicated than we expected.

What a single score gets wrong

The problem with a 1-to-10 mood score is not the number — it's what you do with it. In isolation, today's score tells you almost nothing. What matters clinically is pattern: the direction of change over time, the variability between days (which turns out to be a stronger predictor of anxiety than the average level), the relationship between mood and behaviours like sleep and exercise, and the gap between what someone reports and how they're functioning.

A person who scores 4 today and has been at 4 for three weeks is in a very different situation from a person who scored 4 today but was at 7 last week. The same score, the same immediate experience, but completely different clinical pictures.

The 11 signals we use instead

After reviewing the clinical literature and consulting our advisory board, we identified 11 data signals that, in combination, produce a significantly richer picture of someone's mental state than a single mood score. These include: the mood score itself, mood variability over the past seven days, sleep quality and duration, exercise and movement, social activity level, completion rate of therapeutic exercises, journal sentiment (analysed with explicit user consent), self-reported energy levels, appetite, concentration, and — critically — the trend direction over time.

These signals are weighted differently depending on what we're trying to understand. For anxiety, variability and sleep quality are the strongest predictors. For depression, trend direction and activity level dominate. For overall wellbeing, a composite model performs best.

The debate about transparency vs. precision

The most heated debate we had internally was about how much to explain to users. Our data science team initially built a model that produced a single 'wellbeing score' — a composite index that outperformed any individual signal in predicting clinical outcomes. It was more accurate. It was also a black box.

Our clinical advisors pushed back hard. 'If someone sees a number go down and doesn't understand why, that's not useful — it's anxiety-inducing,' said Dr. Kwame Osei, who sits on our advisory board. 'The insight only has value if the person can act on it.'

We landed on a deliberate trade-off: slightly less precise, but fully explained. Every mood insight in YouMindo tells you not just what the model has noticed, but why — in plain language, sourced to specific inputs. 'Your mood variability has been higher than usual this week — often linked to disrupted sleep. Your sleep scores dropped on Tuesday and Wednesday.' That's more useful than a number going from 7.4 to 6.8.

What we're still getting wrong

The honest version of this post has to acknowledge the limits. Our model is trained primarily on data from users who are already engaged enough to fill in daily check-ins — which is not the same population as people who most need support. People in acute distress often stop tracking. That's a fundamental sampling problem we haven't solved.

We're also aware that mood tracking itself can become a source of anxiety for some users — a phenomenon well-documented in the research literature. We've added deliberate friction: a 'take a break from tracking' option that doesn't penalise engagement scores, and a clinical review trigger if tracking frequency suddenly drops.

The goal was never to build a perfect model. It was to build one that's useful, honest about its limits, and genuinely grounded in the science. We think we're closer to that now than we were in 2022. We're also certain we'll be rebuilding it again.