From Study to Recommendation

In the health sciences alone, over 200,000 papers are published every year. For years I have worked on the harder part: turning them into something the people who have to decide can actually use. A finding sits in a journal, and nobody carries it across into a form that holds up under a real project.

I built a system for that translation. It takes unstructured, often contradictory research material and turns it into recommendations that are parametrised and verifiable: a specific number, tied to a specific source, valid under stated conditions. The system runs as two graphics here: a five-step pipeline from raw study to finished recommendation, and a five-layer structure that classifies every data point it holds.

The examples come from neuroarchitecture because that is where I started. The principle is industry-agnostic. The same system works for food safety or site evaluation. The structure does not change, only the content.

From study to recommendation

The first step is the raw material itself: individual studies, unstructured and often contradictory, more of them than any one person can hold in view. The second step is extraction. Each study gets broken down into what was measured, on whom, with what result and how sound the study design was. The third step is ontology: the extracted data points get placed into a system of relationships, showing what belongs together, what contradicts what and what depends on what. The fourth step, context filter, feeds in a concrete project. An open-plan office with 200 knowledge workers, neurodiversity to be considered, defines the room type, the users and the requirements, and the system filters the knowledge base against that definition. The fifth step is the recommendation itself, carrying its own evidence quality, source and validity conditions.

One example runs through the whole pipeline. Heschong 1999, a study on daylight in schools, analysed the test records of more than 21,000 students across three districts and found that pupils in the most daylit classrooms progressed 20 percent faster on maths tests and 26 percent faster on reading tests over one school year than pupils in the least daylit ones. The study was correlational and named no mechanism; the circadian pathway often proposed for such effects runs from the ipRGC photoreceptors through the SCN to cortisol. Fed into the office context above, the system produces a recommendation of 300 to 500 lux horizontal illuminance, melanopic EDI above 250, between 09:00 and 12:00. The evidence line shown with that recommendation here (a pooled sample size, a Cohen’s d, a GRADE rating) is an illustrative example of the output format, not a figure drawn from the Heschong study, which reported neither an effect size nor a GRADE grade.

How the system thinks

Every data point in the system carries five tags. The values in the examples below are illustrative, chosen to show the format rather than to report measured findings. Context describes the room, its users and its conditions: an open-plan office, knowledge workers aged 30 to 55, a high proportion of focused individual work, neurodiversity to be considered, located in Munich, south-facing. Intervention describes what changes, always in parametrised and measurable form: daylight with melanopic EDI above 250 at the workstation between 09:00 and 12:00, reverberation time RT60 below 0.6 seconds, CO2 concentration below 800 ppm. Both feed into effect, the outcome itself: sustained attention improves (Cohen’s d=0.45), stress recovery accelerates (cortisol reduction in 10 minutes), sick days decrease by 15 percent.

Two further tags qualify the effect rather than restate it. Evidence states how solid the finding is: GRADE moderate, basis 3 RCTs and 1 cohort study, downgraded for indirectness (lab studies, not field data), no downgrade for inconsistency (the studies agree). Condition states what changes the effect: the daylight effect on attention is stronger in morning chronotypes; for evening chronotypes the optimal time window shifts; people over 65 need higher lux values because of reduced pupil aperture.

The examples come from neuroarchitecture because that is where I started. Food safety and site selection run through the same five steps and the same five tags. Only the studies feeding the system change.