Stop imputing every gap: patient knowledge graphs hold their clustering quality across three missingness mechanisms

Patient stratification no longer has to fill every gap before reducing dimensions. In specialist records, missingness often depends on the missing value itself: autoantibody tests are ordered only when a diagnosis is suspected, and symptoms are recorded mainly when present, so imputation treats an untested antibody and a confirmed negative alike. The usual route imputes medians, then reduces with UMAP, then clusters; under value-dependent missingness that pipeline gave the lowest clustering quality. Across 352 analysable patients in a simulation comparison, the knowledge graph representation had a floor NMI of 0.395 and a range of 0.061 across the three mechanisms, while imputation plus UMAP reached a floor of 0.083 with a range of 0.265.

The construction makes absence carry no information. Each patient is a node, and observed diagnoses, antibody results and symptoms become typed edges to concept nodes; a test with no recorded result adds no edge, so untested is never written as negative. TransE then maps nodes and relations into a vector space, treating each relation as a translation, and k-means clusters the patient vectors. The graph holds 352 patient nodes, 49 concept nodes and 1,408 triples, clustered with k = 6; all seven laboratory biomarkers excluded from graph construction differed significantly across the six clusters after Benjamini-Hochberg correction (q < 0.01).

The findings rest on 359 consecutive patients from a single specialist allergy and clinical immunology unit in Italy, of whom 352 had diagnostic labels and entered the graph experiments. Clustering quality is measured as normalised mutual information against a macro-clinical area label, and the six clusters are externally checked against seven laboratory biomarkers not used to build the graph. The simulation uses discretised laboratory values as a proxy for the missingness structure of antibody and symptom data, covering random, conditional and value-dependent missingness.

The data come from a single specialist centre, and external validity requires multi-centre replication. The primary validation metric shares diagnostic tag information with graph construction, a partial circularity the authors mitigate with independent laboratory biomarker validation. The three simulated mechanisms do not exhaust clinical missingness, TransE and the 64-dimensional embedding were not systematically optimised, and statistical comparisons over five seeds have limited power.