An increasing number of lung nodules are found in asymptomatic patients. Physicians are limited in their ability to predict which nodules are malignant despite a variety of imaging techniques. Tissue diagnosis is the only reliable way to evaluate lung nodules, exposing the patient to variable risks depending on the method of biopsy.
We retrospectively reviewed our previous 100 cases of lung nodules with confirmed tissue diagnosis, which presented to our practice over the last 2 years. Fifty factors including, but not limited to, a patient’s past medical history, co-morbidities, pulmonary function tests, nodule size, PET scan values, laboratory values were recorded. These were used to build a linear classifier (a support vector machine). It locked onto 109 features out of the maximum total of ∼1300 numbers generated for each patient (50 original numbers and then all-pairs). The classifier was trained on 80% of the patients and tested on the remaining 20%. The support vector machine finds the largest margin plane separating the malignant from the benign training examples.
This model was able to predict the nodule being benign or malignant correctly 91% of the time.
Using Bayesian Networks and support vector machines, new methods to evaluate unknown outcomes are possible. These new tools of prediction analysis may assist physicians in determining an individual patient’s personal risk of a nodule being benign or malignant. Previous statistical analysis using linear relationships does not allow individual risk analysis.
Treatment decisions for the patient with an indeterminate pulmonary nodule can be guided more effectively with these new methods of prediction analysis. The risk of invasive diagnostic procedures may be reduced by recommending intervention to those who are more likely to need them. These methods have implications for other areas in medicine where prediction analysis is required.
R.C. Ashton, None.