Combining Artificial Neural Networks and Transrectal Ultrasound in the Diagnosis of Prostate Cancer

Drs. Porter and Crawford carefully assess the role of artificial neural networks (ANNs) as predictive models of outcomes for initial prostatic biopsies performed in conjunction with transrectal ultrasound (TRUS). Obviously, the treatment of prostate cancer rests on establishing the diagnosis via biopsy, and TRUS-guided core biopsies have been the standard of care since Hodge et al reported the superiority of this technique in 1989.[1] Nevertheless, of the more than 1 million TRUS-guided biopsies performed annually in the United States, only one-fifth result in the detection of cancer.[2] Thus, to overstate the obvious, a predictor of biopsy outcomes is not only desirable, but is of utmost importance, when costs, anxiety, and time have to be factored into a procedure with a yield of only approximately 20%. When to Initially Biopsy Patients
Prior to the advent of prostatespecific antigen (PSA) assays, urologists relied almost exclusively on the results of a digital rectal examination (DRE) as the indication for prostate biopsy. Having to rely on the discernment of a nodule as the sole criterion for biopsy frequently resulted in diagnosis when the patient already had metastatic disease. With the discovery of PSA and the use of TRUS, the rate at which clinically localized prostate cancer is detected rose dramatically. With the variability in the size of the prostate and the ongoing debate as to when to perform an initial biopsy, urologists need to make a decision about the initial biopsy practically on a daily basis. To date, the parameters of PSA, DRE, age, and ultrasound for initial biopsy are constantly being refined with other prebiopsy markers, eg, percent- free PSA, complex PSA, PSA velocity, and PSA density. The proliferation of new markers, while potentially improving the prediction of biopsy outcomes, also complicates clinical judgment. Clinical judgment is indispensable, but as Kattan pointed out, that judgment is subject to inherent bias.[3] Given this background, one should consider the value of predictive models to guide the urologist in deciding when to perform the initial biopsy. In an ideal predictive model, this dilemma would be solved. However, as the authors point out, there are at least six validated predictive ANN models that use prebiopsy parameters to predict prostate biopsy outcomes, however these various ANNs are not uniform with regard to input variables.[ 4] In additon, there remains the complexity of constantly changing practice patterns that affect the input variables-eg, the number of biopsies are not uniform in the various models. Moreover, in reports of one of the original ANNs (the work of Snow et al), the authors did not describe their results using the receiver- operator characteristics (ROC) curve, a technique that is now used almost universally.[4] Drs. Porter and Crawford recognize these problems, and in discussing the European Djavan model (in which two biopsies were performed if the first biopsy was negative), they state, "caution should be exercised in applying this result to other contemporary series."[5] Conclusions
On balance, it appears that ANNs are as accurate as their logistic regression counterparts. As the authors point out, because biologic systems are not linear, ANNs have the theorectical advantage of being more refined predictive models. Although the accuracy of the predictive models ranges from 0.75 to 0.91, most of them are of limited value due to the fact that their application is limited to a distinct subset of patients and usually will not transcend a particular clinical situation. In essence, "one size does not fit all." Porter and Crawford go to great lengths to make it clear that when counseling patients regarding biopsy, a model with a predictive accuracy of over 0.75 may, nonetheless, be preferable to no model. I feel that the ongoing challenge is the need for more cooperation among ANN investigators. Although the current models do not fit all situations, funding to bring together experts in these fields (to develop truly userfriendly models) must be sought in grant applications and at all levels of government. Ideally, we need to develop models that transcend practice patterns in different parts of the world. The authors also clearly state the need for individual practitioners to have easy and immediate access to these models. The authors are to be commended for succinctly stating the issues surrounding ANNs. They explain what must be done to get a handle on ongoing advances in medicine-eg, immediate Internet access-which are frequently not translated into immediate adjuncts in daily clinical practice. The article focuses on the positive predictive value of ANNs for an initial biopsy. With better models, the number of initial biopsies should decrease. We will still need to discern the necessity of subsequent biopsies after an initial negative biopsy-eg, in the case of a rising PSA. For this group of patients, in particular, and all patients in general, we need better predictive models.