- Meeting abstract
- Open Access
The use of neural network algorithms for the classification of undifferentiated arthritis by synovial histopathology
© The Author(s) 2003
- Received: 14 January 2003
- Published: 24 February 2003
- Neural Network
- Decision Tree
- Positive Predictive Value
- Definite Diagnosis
We indicated recently that synovial histopathology can be used for early diagnosis of undifferentiated arthritis (UA), with multiparameter models such as decision trees being superior to single parameters. The present aim is to build a prediction method based on a neural network approach and to compare this to the decision tree algorithm.
Synovial histopathology (including14 parameters) was performed on 120 synovia of patients with knee synovitis presenting for diagnostic workout. In 67 patients, a definite diagnosis was made at the time of arthroscopy (32 RA, 22 SpA, 13 other diseases); this cohort was used as learning file for the algorithms. The remaining 53 patients had real undifferentiated arthritis (UA) at presentation and a definite diagnosis was obtained by clinical follow-up (19 RA, 21 SpA, 13 other diseases) 6 months later; this cohort was used as test file to validate the models. A self-organizing neural network was developed using modified Kohonen mapping in combination with a case-based learning evaluation criterion.
Classifying all 53 UA, the global positive predictive value (PPV) of the neural network was 64.2%, including a PPV of 78.9% for RA, 57.9% for SpA, and 63.2% for other diseases. Using the decision tree model on the same cohort, the global PPV was 69.8%, with a PPV of 66.7% for RA, 70.0 % for SpA, and 77.8% for other diseases. Since it is clinically more important to classify a subset of UA with a high PPV (> 80%) than to classify all patients (resulting in a lower overall PPV), a distance criterion was placed on top of the case-based learning evaluation algorithm, resulting in the classification of 50.9% of the samples with a PPV of 81.5%. Using the same PPV requirement for the decision tree, 79.2% of the samples were classified with a PPV of 81.0%.
Although the global performance of the neural network was slightly inferior to the decision tree model, the neural network was superior for the classification of RA. These results warrant further research in the neural network approach, especially for classification systems using a high number of parameters.