How a map is made
Search PubMed for the analyte, sorted by relevance, and take up to
400 abstracts. Run scispaCy's biomedical named-entity recognizer
over them to pull out every disease and chemical mentioned. Turn
each term into a 200-dimension word vector trained on PubMed text,
then use t-SNE to press those vectors down to two dimensions so
they can be drawn.
Nothing runs when you load this page. The maps are computed ahead
of time with a fixed random seed, so a given map looks the same on
every visit.
Why you can't compare two maps
t-SNE coordinates carry no meaning between runs. Position on the
canvas is arbitrary, and the distance between two clusters says
nothing. Only local neighbourhoods carry information: which terms sit next to
which, and only within a single map.
Switching analytes gives you a different question, not a
comparison. The stability score below each map measures how well
those neighbourhoods survive re-running with a different seed.
It also falls as maps get larger, so it can't rank one analyte
against another either.
t-SNE will cluster noise
Give t-SNE structureless data and it still returns a picture with
clean, well-separated blobs. So "the plot has clusters" is not
evidence of anything. That's what the null model is for: it shuffles
each dimension independently, destroying the relationships between
terms while preserving every other property of the data.
Toggle it on. If you can't tell the real map from the null, the
real map isn't showing you anything.
The literature isn't always about the strip
Queries search each analyte broadly rather than restricting to the
urine or saliva finding. That surfaces associations a narrower
search would miss, but it means several maps describe the blood
version of a measurement: the glucose literature is overwhelmingly
blood glucose, leukocytes is largely haematology rather than
urinary leukocyte esterase.
Where that applies, it's stated on the analyte itself rather than
left to this paragraph.
Most of every map is one paper's vocabulary
Between 72% and 88% of the terms in these maps appear in only a
single abstract. Those points are drawn faintly and left unlabelled.
They are not noise exactly. A rare term may be precisely the
under-studied association worth chasing. But a term used by
one paper can't distinguish a real signal from a passing mention.
What was left out
Clustering was computed and then dropped from this page. K-means on
these vectors produced a median silhouette of 0.10 across 22 maps,
which is barely above no structure at all, with the largest cluster
routinely swallowing half the terms. Showing those groupings as
findings would have been the most misleading thing here.
Spelling and lemma variants of one word are merged
(haematuria into
hematuria,
diabete into
diabetes). Distinct concepts are
not: diabetes,
t2dm and
gdm remain separate points, which
means one condition can be scattered across several. Resolving that
properly needs concept linking, not string matching.