Machine Learning-based Online Survival Prediction Tool for Adrenocortical Carcinoma

Saygili Emre Sedar; Elhassan Yasir S; Alessandro Prete; Juliane Lippert; Barbara Altieri; Ronchi Cristina L

doi:10.1530/endoabs.99.EP14

EP14

Prev Next

Section Contents Cite

Endocrine Abstracts (2024) 99 EP14 | DOI: 10.1530/endoabs.99.EP14

ECE2024 Eposter Presentations Endocrine-Related Cancer (90 abstracts)

Machine Learning-based Online Survival Prediction Tool for Adrenocortical Carcinoma

Emre Sedar Saygili ¹ , Yasir S Elhassan ² , Alessandro Prete ² , Juliane Lippert ³ , Barbara Altieri ⁴ & Cristina L Ronchi ²

Views

Author affiliations

¹Division of Endocrinology and Metabolism, Department of Internal Medicine, Faculty of Medicine, Canakkale Onsekiz Mart University, Çanakkale, Turkey; ²Institute of Metabolism and Systems Research, University of Birmingham, United Kingdom; ³Institute of Human Genetics, University of Wuerzburg, Germany; ⁴Division of Endocrinology and Diabetes, University Hospital of Wuerzburg, Germany

Background: Adrenocortical carcinoma (ACC) is a rare endocrine cancer. We aimed to develop machine learning (ML) models for predicting clinical outcomes of patients with ACC and deploy them as a web-based decision support tool.

Methods: The S-GRAS dataset¹ was used as a training cohort (n=942), while the COMBI dataset² and new patients were used as a validation cohort (n=220). We used S-GRAS parameters previously described¹ for ML models. The PyCaret 3.1.0 ML library in Python was used to create models. The F1 score is calculated as the harmonic mean of sensitivity and precision. We compared sixteen ML models and chose the best by F1 score. Clinical outcomes were defined as 5-year overall mortality (OM), 1-year disease progression (DP), and 3-year DP, respectively.

Results: The study has 579, 968, and 858 patients’ data for clinical outcomes, respectively. The study’s 5-year OM, 1-year DP, and 3-year DP rates were 55.1%, 39.7%, and 67.6%, respectively (training+validation cohorts). Quadratic Discriminant Analysis (QDA), Light Gradient Boosting Machine (LGBM), and AdaBoost Classifier (ABC) were the best models for predicting clinical outcomes. The F1 scores of the best ML models for the training cohort were 0.79 for OM, 0.63 for 1-year DP, and 0.83 for 3-year DP; while for the validation cohort were 0.72, 0.60, and 0.83, respectively. Sensitivity and specificity for 5-year OM were 77% and 77% in the training cohort and 65%, and 81% in the validation cohort, respectively. Streamlit in Python is used for deploying the models as a website (https://acc-survival.streamlit.app).

Table 1. Performance of the best ML models
Outcomes	Best model	Cohort	Accuracy	AUC	Sensitivity	Precision	F1	Specificity
OM	QDA	Training	0.77	0.85	0.77	0.80	0.79	0.77
	Validation	0.73	0.79	0.65	0.80	0.72	0.81
1-year DP	LGBM	Training	0.74	0.78	0.57	0.70	0.63	0.84
	Validation	0.68	0.74	0.53	0.71	0.60	0.81
3-year DP	ABC	Training	0.76	0.79	0.88	0.83	0.83	0.50
	Validation	0.77	0.87	0.79	0.83	0.83	0.71

Conclusion: We could demonstrate that S-GRAS parameters can predict OM and disease progression with an AUC range of 0.74-0.87. To the best of our knowledge, this is the first ML-based online survival tool for ACC. This app instantly gives the probability of outcomes for patients with ACC based on S-GRAS parameters after resection. Medical professionals could use it in clinical practice to drive personalised management decisions.

References: 1. Elhassan YS, et al. Eur J Endocrinol.2021;186¹:25-36.

2. Lippert J, et al. Eur J Endocrinol. 2023;189(2):262-270.