Artificial neural network model for the assessment of lymph node spread in patients with clinically localized prostate cancer

• Published in: Urology (Ridgewood, N.J.) Vol. 57; no. 3; pp. 481 - 485
• Main Authors: Batuello, Joseph T, Gamito, Eduard J, Crawford, E.David, Han, Misop, Partin, Alan W, McLeod, David G, O’Donnell, Colin
• Format: Journal Article
• Language: English
• Published: New York, NY Elsevier Inc 01.03.2001; Elsevier Science
• Subjects: Biological and medical sciences; False Negative Reactions; Humans; Lymph Node Excision; Lymphatic Metastasis; Male; Medical sciences; Neoplasm Staging - methods; Nephrology. Urinary tract diseases; Neural Networks (Computer); Pelvis; Prostatectomy; Prostatic Neoplasms - pathology; Prostatic Neoplasms - surgery; ROC Curve; Sensitivity and Specificity; Tumors of the urinary system; Urinary tract. Prostate gland; Human; Evaluation; Focal lesion; Urinary system disease; Prognosis; Prostate disease; Lymph node; Stage classification; Male; Metastasis; Neural network; Malignant tumor; Modeling; Decision aid; Male genital diseases; Artificial intelligence; Prostate
• ISSN: 0090-4295; 1527-9995
• DOI: 10.1016/S0090-4295(00)01039-6

Objectives. To develop an artificial neural network (ANN) model to predict lymph node (LN) spread in men with clinically localized prostate cancer and to describe a clinically useful method for interpreting the ANN’s output scores. Methods. A simple, feed-forward ANN was trained and validated using clinical and pathologic data from two institutions (n = 6135 and n = 319). The clinical stage, biopsy Gleason sum, and prostate-specific antigen level were the input parameters and the presence or absence of LN spread was the output parameter. Patients with similar ANN outputs were grouped and assumed to be part of a cohort. The prevalence of LN spread for each of these patient cohorts was plotted against the range of ANN outputs to create a risk curve. Results. The area under the receiver operating characteristic curve for the first and second validation data sets was 0.81 and 0.77, respectively. At an ANN output cutoff of 0.3, the sensitivity achieved for each validation set was 63.8% and 44.4%; the specificity was 81.5% and 81.3%; the positive predictive value was 13.6% and 6.5%; and the negative predictive value was 98.0% and 98.1%, respectively. The risk curve showed a nearly linear increase (best fit R 2 = 0.972) in the prevalence of LN spread with increases in raw ANN output. Conclusions. The ANN’s performance on the two validation data sets suggests a role for ANNs in the accurate clinical staging of patients with prostate cancer. The risk curve provides a clinically useful tool that can be used to give patients a realistic assessment of their risk of LN spread.