Overall Survival Prediction for Patients Diagnosed with Acute Myeloid Leukemia

Evaluating the Incremental Value of Cytogenetic & Molecular Features

Author: Dorjderem Namsraijav (3035962362)
Supervisor: Asst. Prof. Liangqiong Qu
Department of Computing and Data Science, The University of Hong Kong
April 2026

Why This Matters

Acute myeloid leukemia (AML) is a biologically diverse blood cancer with highly variable patient outcomes. Accurate risk stratification at diagnosis is essential for guiding treatment decisions – low‑risk patients may benefit from supportive therapy alone, while high‑risk candidates could be prioritised for intensive approaches like stem cell transplantation. However, advanced cytogenetic and molecular tests remain unavailable in many community hospitals, creating a gap between recommended diagnostics and everyday practice. This project quantifies how much extra prognostic value these specialised tests add beyond routine clinical variables, helping clinicians and hospitals decide where to invest resources.

Data & Models

• Dataset: 3,323 AML patients from Institut Gustave Roussy (France).
• Clinical features (8): age, sex, blood counts (WBC, ANC, monocytes, haemoglobin, platelets), bone marrow blast percentage, treatment centre.
• Cytogenetics: raw karyotype strings engineered into structured features (complexity, specific abnormalities, chromosome‑level events).
• Molecular: 10,935 somatic mutations aggregated per patient – mutation burden, effect types, VAF statistics, gene/chromosome spread.
• Models: Cox Proportional Hazards (L2‑penalised), XGBoost (Cox objective), Random Survival Forest – all evaluated with Uno’s IPCW concordance index (c‑index).
• Validation: Nested 5‑fold cross‑validation, repeated twice. Statistical significance tested with the Wilcoxon signed‑rank test.

Key Results

Adding either cytogenetic or molecular data already gives a statistically significant boost. The full combined model achieves the highest discriminative performance, indicating that these specialised tests deliver complementary information not captured by routine labs alone.

1. Feature Importance (XGBoost)

Feature importance analysis confirms that the extra diagnostic layers provide the most influential predictors. Only five clinical features appear among the top 20, highlighting the added value of deeper testing.

2. Correlation Among Clinical Features

3. Clinical Tests vs. Reference Ranges

4. Normal Test Results & Survival

Patients with a higher number of normal laboratory values consistently show longer average survival. This simple derived feature contributes to the clinical‑only model’s baseline performance.

5. Fold‑level C‑index Consistency

Statistical Significance

All pairwise comparisons against the clinical‑only baseline were evaluated with the dependent Wilcoxon signed‑rank test across 2×5 repeated folds. Every extended feature set showed a statistically significant improvement (p = 0.002), confirming that the observed gains are not due to chance.

Conclusion & Clinical Relevance

This work demonstrates that cytogenetic and molecular data significantly improve survival prediction in AML beyond what can be achieved with standard clinical variables alone. The combined model reaches a c‑index of 0.73, representing a meaningful gain in the ability to correctly order patients by risk. While routine labs already provide a solid prognostic baseline (c‑index ~0.68), the extra 5‑point improvement could be decisive when stratifying patients for trials or intensive treatments.

From a hospital operations perspective, the results justify investments in specialised testing infrastructure or rapid referral pathways, especially for centres that currently rely solely on clinical variables. Nevertheless, external validation and calibration analyses are needed before the model can be deployed in real‑world clinical workflows.

Project Resources

Full code and documentation:
https://github.com/Dorjderem2002/fyp

Data provided by:
Institut Gustave Roussy (Challenge Data)