Untangling profiles of post-thrombotic syndrome using unsupervised machine learning

Postthrombotic syndrome (PTS) is a chronic condition that can develop after deep vein thrombosis (DVT) and is diagnosed using the Villalta scale. This study applied unsupervised machine learning to investigate the heterogeneity of PTS among patients and within the Villalta scale. In 818 patients from the IDEAL DVT study, clustering identified 4 clinical profiles: (1) younger patients with provoked DVT, (2) women with joint pain, (3) men with isolated popliteal DVT, and (4) older men with diabetes and femoral vein involvement. Clustering of Villalta items revealed a distinction between signs and symptoms. Sign scores increased with older age, male sex, higher body mass index (BMI), and DVT extent, whereas symptom scores increased with younger age, female sex, higher BMI, and provoked DVT. Residual venous obstruction was significantly associated with the sign score (odds ratio, 1.18 per point) but not the symptom score. Quality of life was related to the symptom score more than the sign score. At 6 months, sign and symptom scores differed significantly across profiles, especially between profile 1 and 4, because the former had most symptoms (41% vs 21% ≥ 3; P < .001), whereas the latter had most signs (18% vs 34% ≥ 3; P = .004]). After 2 years, symptoms decreased in profile 1 but increased in profile 4. Other profiles showed intermediate scores over time. These findings suggest that reappraising the PTS scoring system to distinguish its dimensions would enable more personalized risk prediction and prevention.