B14
Development of a “composite serum biomarker” model to predict presence of colorectal cancer using artificial neural networks: an example with IGF-related peptides
Lee Lancashire, Ivona Baricevic, Darren Roberts, Caroline Dive, Andrew Renehan
Paterson Institute for Cancer Research, Manchester, UK
Background
The development of “composite serum biomarkers” that predict the presence of cancer using readily analysed, inexpensive peptide assays is appealing. This study investigated multi-dimensional representations of relationships within a panel of serum insulin-like growth factor (IGF) peptides with age, sex, and body mass index (BMI) using artificial neural networks (ANNs) to derive predictive models for colorectal cancer (CRC).
Method
ANNs were developed sequentially via training, testing and validation. Healthy individuals controls (n = 722, aged 20 to 85 years; M:F, 384:338), CRC patients (n = 74) and acromegaly patients (n = 52, as “positive” discriminators) were analysed in a case-control design. Predictability was expressed as sensitivity, specificity and AUC derived from combinational ROC curves.
Results
Curved decision surfaces were generated for various combinations of analytes through 100 iterations of cross-validation. Using stepwise forward selection, the optimum combination discriminating CRC versus normal was IGFBP-2, IGF-I, IGFBP-3 with age and BMI, with sensitivity of 86%; specificity, 95% and AUC = 0.96 (95% CIs: 0.92, 1.00). For acromegaly versus normal, the optimum variable combination was IGFBP-3, BMI, IGF-I, IGFBP-2, IGF-II and age (sensitivity 96%; specificity, 96%, and AUC = 0.99 (95% CIs: 0.99, 1.00). This model outperformed a stepwise logistic regression approach significantly.
Conclusion
ANNs identify data patterns by knowledge acquisition through an iterative learning process that allows development of non-linear models. This study demonstrated the power of this innovative approach to develop a “composite biomarker” that predicted the presence of CRC.
