TL;DR
This paper investigates whether artificial neural networks (ANNs) can improve breast cancer survival prediction accuracy beyond the established TNM staging system. The study found that several ANNs and logistic regression models significantly outperformed TNM, with backpropagation neural networks achieving the highest accuracy.
What Problem Does This Paper Address?
The TNM staging system, a cornerstone of breast cancer prognosis since the 1960s, has limitations. Its accuracy is only about 44%, and it cannot easily incorporate new prognostic factors that have been identified over time. This proliferation of factors has led to clinical confusion. The authors aim to explore whether computerized prognostic systems, particularly ANNs, can integrate these new factors with TNM variables to enhance predictive accuracy. They hypothesize that ANNs, being non-linear models, might better capture complex interactions and non-monotonic relationships between prognostic factors, which traditional linear models might miss.
Methods
The study compares the predictive accuracy of several models for five-year breast cancer-specific survival. The models evaluated include:
- pTNM staging system: The standard pathological staging system.
- Principal Component Analysis (PCA): A data reduction technique.
- Classification and Regression Trees (CART): Two versions, a pruned 9-node tree and a shrunk 13.7-effective-node tree.
- Logistic Regression: Performed stepwise with cubic splines for age to avoid assuming linearity.
- Artificial Neural Networks (ANNs):
- Fuzzy ARTMAP: An adaptive resonance system.
- Cascade Correlation ANN: A type of feedforward network.
- Conjugate Gradient Descent ANN: A gradient-based optimization method.
- Probabilistic Neural Network (PNN): Based on Bayesian probability.
- Backpropagation ANN: A widely used gradient descent algorithm for training multilayer perceptrons.
Accuracy was measured using the area under the receiver operating characteristic (Az) curve, a non-parametric measure of discrimination. The data was randomly split into training, hold-out, and testing sets.
Datasets
The study utilizes the Patient Care Evaluation (PCE) data set, collected by the Commission on Cancer of the American College of Surgeons (ACS). This dataset contains cancer information from hospital tumor registries in the United States, specifically the first 25 cases of breast cancer seen at each institution in 1983, with follow-up information up to the date of the request. The dataset provides up to eight years of follow-up. For this analysis, cases with missing data or censored before five years were excluded. The data was divided into training (3,100 cases), hold-out (2,069 cases), and testing (3,102 cases) subsets.
Key Insights and Ideas
- ANNs as a powerful tool for complex medical data: The study strongly suggests that ANNs can effectively handle the complexity of prognostic factors in breast cancer, potentially outperforming traditional models when non-linear relationships or interactions are present.
- Limitations of TNM: The TNM system's fixed nature and limited accuracy highlight the need for more adaptable and accurate predictive models in clinical practice.
- Importance of model selection: Different ANN architectures and traditional models exhibit varying levels of accuracy, emphasizing that the choice of model is crucial for optimal prediction.
Key Results
- The pTNM staging system achieved an accuracy (Az) of .720.
- Principal Components Analysis was less accurate, with an Az of .714.
- CART models showed accuracies of .753 (pruned) and .762 (shrunk).
- Logistic regression with cubic splines achieved an accuracy of .776.
- Among the ANNs:
- Fuzzy ARTMAP had the lowest accuracy at .738 and was computationally intensive.
- Cascade correlation and conjugate gradient descent showed potential, with accuracies of .761 and .774, respectively.
- Probabilistic Neural Network (PNN) achieved an accuracy of .777, but had significant storage requirements.
- Backpropagation ANN demonstrated the highest accuracy at .784.
Overall, several computerized models, including logistic regression and various ANNs, were found to be significantly more accurate than the TNM staging system for predicting five-year breast cancer-specific survival.
Limitations and Future Work
- Computational Intensity: Fuzzy ARTMAP was noted as being too computationally intensive to be a practical model.
- Storage Requirements: PNN, despite its good accuracy, has significant storage requirements.
- Focus on previously identified factors: The study acknowledges that the analyses rely on previously collected prognostic factors that were already identified as significant in generalized linear models. The paper's objective was to show ANNs could perform as well as traditional models with these factors, rather than discovering entirely new ones.
- Future Directions: The authors suggest that ANNs are likely to outperform conventional models when unanticipated nonmonotonic factors or complex interactions exist, implying further research into these scenarios would be valuable.
Summary updated