A single score does not describe a trajectory
Two People can have the same average across a month and arrive there through very different sequences. One may remain close to a personal baseline. Another may move sharply between observations. A third may shift after an event and return gradually. The average is still correct, but it is not the whole pattern.
Affective Dynamics is the study of how affective states change across time. The field examines trajectories and regularities in repeated observations, including average level, variability, successive change and temporal dependence. These are descriptions of a measured series. They do not explain why a change happened and they do not make a Diagnosis.
The established measures answer different questions
- Average level. The within-person mean describes the centre of the available observations. It is sensitive to the Instrument, scale, window and missing observations that produced it.
- Variability. Within-person variance or standard deviation describes dispersion around that centre. It does not preserve the order in which observations occurred.
- Successive change. Successive-difference measures retain temporal order by asking how far adjacent observations move. Unequal intervals require explicit treatment because a change over two hours is not the same process as the same change over two days.
- Inertia. Lag-one autocorrelation asks how strongly one observation predicts the next. It becomes difficult to interpret when timing is irregular, the series is short or the measure itself is unstable.
- Recovery. A return toward a reference band can be described only within the available observation window. If the return is not observed, the honest result is unresolved rather than an invented duration.
Common summaries include the within-person mean, variance, mean squared successive difference and lag-one autocorrelation. Each keeps different information and each inherits the limits of the sampling design.
What Heyrafiki is testing
We are testing this approach with example check-in sequences and a Person's available usual range. The original check-ins remain distinct from the summary and from any interpretation a Practitioner chooses to record.
- Personal continuity before population labels. The first comparison is with the Person's available baseline window, not a claim about what a population should look like.
- Time remains part of the evidence. Irregular intervals, gaps and the end of the observation window stay visible. Missing observations are not silently filled in.
- Uncertainty can stop a measure. A short, constant or poorly timed series returns an explicit not-interpretable state instead of a confident-looking number.
- Review remains human. The pattern may support a conversation. It cannot become an interpretation in the Clinical Record until an authorised Practitioner reviews and approves it.
More dynamics do not automatically mean more knowledge
The scientific literature gives a serious reason for restraint. A large meta-analysis found relationships between several short-term dynamic patterns and well-being. Later work across 15 studies found that more complex dynamic measures added little beyond mean levels and variance for the outcomes tested. These findings are not contradictions to hide. They define the validation question: which measure adds reliable, decision-useful information for a declared purpose, population and timescale?
Reliability is also not guaranteed by having many observations. Sampling error, measurement invariance, response burden, missingness and the spacing of prompts can change what an index means. A technically valid equation can still produce a weak clinical signal when the measurement design is weak.
This approach is not a Diagnostic model, a deterioration alarm or a clinical conclusion. Events that happen close together do not prove that one caused the other. Screening and longitudinal context support a Clinician's judgement; they do not replace it.
The evidence required before clinical use
- Measurement validity. The construct, Instrument, language, population, responder role and repeated-use properties must be appropriate for the intended setting.
- Sampling robustness. Benchmarks must test sparse series, irregular spacing, missed prompts, changing baselines, extreme variability and observation windows that end before recovery.
- Incremental value. A dynamic measure must outperform simpler baselines for its declared decision without rewarding noise or merely restating the mean.
- Clinical usefulness and safety. Practitioners and People must be able to understand the view, contest it and act without mistaking it for a Diagnosis.
- Equity and local validity. Performance must be evaluated across languages, ages, settings and access patterns relevant to the People the system is intended to serve.
- Reproducibility. A declared protocol, example datasets, versioned methods, tests that should fail and the final publication decision must remain inspectable together.
What the evidence supports today
Today, the work is supported by repeatable tests with difficult example time series, clear limits when there is not enough information and a separate step for Practitioner review. That is reproducible engineering evidence. Clinical validation and improved outcomes require prospective protocols, appropriate data and independent review.
