Colposcopic Imaging Carries Most of the Signal: Externally Validated Tri-Modal Integration of Imaging, Vaginal...

Background/Objectives: To determine whether a coagulation-immune index, vaginal microbiome composition, and colposcopic imaging carry non-redundant information about HPV-associated lower genital tract co-infection when integrated within a single learned model. Methods: This retrospective three-center study analyzed 560 women referred for colposcopy: 420 formed the development cohort, and 140 were a withheld external cohort from a third center. All three modalities were sampled within one visit. Participants were classified as HPV-negative, HPV-positive without co-infection, or HPV-positive with co-infection, the latter confirmed by nucleic acid amplification for Chlamydia trachomatis, Neisseria gonorrhoeae, Trichomonas vaginalis, or Mycoplasma genitalium, with overlapping taxa masked a priori. TIM-Fusion combined a ConvNeXt-Tiny image encoder, a feature-tokenizer transformer for the D-dimer-to-lymphocyte ratio, and a perceptron for centered log-ratio abundances through modality-confidence weighting and bottleneck cross-attention. Results: On the external cohort, TIM-Fusion achieved a macro-AUC of 0.92 (95% CI 0.86 to 0.96) and accuracy of 0.90 across the three categories, exceeding the image-only baseline at 0.81, the tabular-only baseline at 0.76, and the best pairwise combination at 0.86. All thirteen comparisons, differences of 0.03 to 0.16, retained significance under Holm correction. Ablation attributed 0.05 (0.01 to 0.10) to bottleneck fusion, the only component with an interval excluding zero; withholding imaging, microbiome, or the ratio cost 0.13, 0.06, and 0.04. Imaging was therefore the dominant contributor, its removal costing three times as much as either of the other modalities. Entering D-dimer and lymphocyte count separately rather than as a ratio gave 0.89 (0.82 to 0.93), so the ratio form is an estimation convenience at this sample size and not a requirement. Co-infection-class AUC was 0.87 (0.75 to 0.94) on 26 events. In a secondary analysis restricted to the 84 HPV-positive external participants, binary discrimination of co-infection was 0.86 (0.76 to 0.93) with sensitivity 0.85 and specificity 0.91, lower than the one-versus-rest figure, which includes HPV-negative women in the negative class. Conclusions: Colposcopic morphology carries most of the discriminatory signal, while vaginal microbial ecology and coagulation-immune status each add information the others do not. Discrimination of co-infection status itself is more modest than the macro-averaged figure suggests, since within the HPV-positive stratum alone it falls to 0.86 on 26 events.