Machine-learning algorithms have been shown to make medical diagnoses as accurate as doctors in the field of cardiology, dermatology, and oncology. 1,2,3By using machine-learning approaches, diagnostic success may be increased whilst human error rate could be reduced. Jocelyn S. Gandelmanfrom Vanderbilt University Medical Center, Nashville, TN, USA, and colleagues published a study ahead of print in Haematologicaassessing machine-learning methods in chronic graft-versus-host disease (cGvHD) patients to identify distinct clusters of phenotypes and survival patterns. 4
The study group investigated whether a machine-learning algorithm could diagnose patients according to organ scores, identify phenotypic subgroups and also stratify survival. Clinical data was used from 339 patients.
- Seven patient groups with contrasting clinical risks were identified
- High-risk cGvHD patients had poor overall survival in comparison with low-risk cGvHD patients: HR = 2.24 (95% CI, 1.36-3.68)
- Clinical applicability:
- Information was converted into a clinical prognostic decision tree
- Groups were selected by the decision tree
- Intermediate- and high-risk patients had inferior overall survival compared to patients in the low-risk group: HR = 2.79, HR = 1.78, respectively
This study shows that machine-learning is promising in terms of identifying biomarkers and stratifying risk in patients with cGvHD. The authors stated that “these results have the potential to be applied to stratify risk in the clinical setting, enhance the current cGvHD classification system, refine inclusion criteria for phase II trials, and guide biomarker discovery for more specific therapeutic targets.” Further studies are required to confirm this data.