Statistical Assessment of Cancer Predictors Using DNA Microarray Data
Author Information
Author(s): Ancona N, Maglietta R, Piepoli A, D'Addabbo A, Cotugno R, Savino M, Liuni S, Carella M, Pesole G, Perri F
Primary Institution: Istituto di Studi sui Sistemi Intelligenti per l'Automazione – CNR
Hypothesis
Is the size of the available data set sufficient to build accurate classifiers for cancer diagnosis?
Conclusion
The study found that statistically significant classifiers for colon cancer diagnosis can be trained with as few as 15 examples.
Supporting Evidence
- Using the whole set of genes, the Weighted Voting Algorithm classifier achieved an error rate of 21% (p = 0.045).
- Regularized Least Squares and Support Vector Machines classifiers achieved error rates of 19% (p = 0.035) and 18% (p = 0.037) respectively with only 15 training examples.
- About 6000 genes were found to be correlated with colon cancer pathology (p < 0.05).
- The best performance of the SVM classifier was an error rate of 11% (p = 0.019) with 35 training examples.
Takeaway
Scientists used data from 25 patients to figure out how to tell if someone has colon cancer using gene information, and they found that they could do it with just a few examples.
Methodology
The study used Leave-K-Out Cross Validation and permutation tests to evaluate classification accuracy and statistical significance.
Potential Biases
Potential selection bias in gene selection methods could affect the generalization of the classifiers.
Limitations
The classifiers may be influenced by noise in the gene expression data.
Participant Demographics
25 patients (14 males; mean age: 60 ± 14 years) with colon cancer.
Statistical Information
P-Value
0.019
Statistical Significance
p<0.05
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website