Monash UniversityLPDP

MASTER OF DATA SCIENCE · INDONESIA

Multimodal Fusion of Non-Contrast CT and Clinical Data for Admission Stroke Severity Classification

A Multisite Evaluation in Indonesian Cohorts

Ananta Tri Wijatmiko 34891919

Supervised by Dr. Vanya Valindria

NCCT visual with coloured lesion regions from the final poster
Original poster visual · lesion regions highlighted in colour
283patients with acute ischemic stroke
3cohorts: RSCM-1, RSCM-2, RSPON
2hospitals in Indonesia
Research focusStroke severity classification at admission

01 / BACKGROUND

Can combined data help explain stroke severity?

quantifies stroke severity but requires a trained assessor. is available in stroke care, while early ischemic changes can be subtle.

This study tests whether imaging and clinical information complement each other in predicting four categories, and whether the benefit persists in a cohort unseen during training.

01

Compare multimodal models with single-modality models.

02

Evaluate cross-cohort generalisation using .

Thesis §1–3, pp. 26–33

02 / KEY FINDINGS

The benefit of fusion depends on the cohort.

HOW TO READ THE RESULTS

Internal performance is an . An advantage in one cohort does not mean the model will perform equally well in another hospital.

03 / LIMITATIONS

Findings that need further validation

  • Small, imbalanced cohorts; two cohorts come from the same hospital.
  • Three shared clinical variables: age, hypertension, and diabetes.
  • One seed (42), three folds, and dependence on lesion annotations.
  • Severe-class recall for feature-level concat: 0.00–0.22, with 2–9 severe patients per cohort.

04 / FUTURE WORK

From annotation to automation

  • Validate with multiple seeds and additional cohorts.
  • Test model-selection strategies with cohort-specific tuning.
  • Evaluate clinically thresholded targets and integrate .
  • Replace annotation-dependent slice selection and feature extraction with automated lesion segmentation.
Thesis §5.8–6, pp. 48–49

GLOSSARY