Systems pathology and predicting outcome after radical prostatectomy.
E. Klein,A. Stephenson,D. Raghavan,R. Dreicer
DOI: https://doi.org/10.1200/JCO.2008.17.1991
IF: 45.3
2008-08-20
Journal of Clinical Oncology
Abstract:A major challenge in clinical oncology is to accurately answer every newly diagnosed patient’s plaintive question, “What are my chances, Doc?” Despite recent advances in understanding clinical, pathologic, and genomic-based prognostic factors, our responses remain imprecise, or—as statisticians might observe—our estimates are surrounded by wide confidence intervals. Although physicians are accustomed to practicing clinical medicine with this degree of uncertainty and have adapted to working within its bounds, most patients are not so accustomed and desire a higher degree of certainty. More accurate estimates of response rates, clinical benefit, and the potential for cure should help patients deal with their diagnosis realistically and would provide better insight into the selection of the most appropriate primary therapy and adjunctive therapies. Attempts at predicting individual outcomes after therapy for localized prostate cancer have a long history, dating to early efforts to correlate preoperative grade, stage and prostate-specific antigen (PSA) with biochemical failure. This was followed by the establishment of formal definitions of low-, intermediate-, and high-risk groups using these parameters. These models were based on easily assessed and widely available clinical data and estimated prognosis and the potential need for more aggressive therapy for patients diagnosed early in the PSA era. Contemporaneously, these attempts were systematized and linked to pathologic outcomes by the development of tables predicting the likelihood of organ-confined disease. These suggested that patients presenting with features that yielded a small chance of cure by surgery might best be treated using another modality. Although these studies were validated informally by daily clinical use and formally by statistical comparisons in additional populations, their predictions were imprecise because of wide variations in outcomes for individual patients within each defined risk category. The next advance in outcome prediction was the development of nomograms, based on Cox regression modeling that combined known predictors of outcome as continuous variables, allowing the assignment of a numerical probability of recurrence for an individual patient rather than a relative risk of recurrence for all members of a defined risk group. An example of the utility of this approach was demonstrated for breast cancer, where a nomogram incorporating standard clinical variables predicting the likelihood of residual disease in the axilla after a positive sentinel lymph node biopsy was substantially more accurate (with an area under the curve of 0.72) than 17 experienced clinicians, who as a group did only slightly better than chance (area under the curve 0.54). Currently there are more than 65 published, externally validated prostate cancer nomograms and other tools that use standard clinical parameters such age, clinical or pathologic stage, grade, percent of cancer on biopsy cores, and PSA or its derivatives to predict various clinical and pathologic outcomes. Although these have achieved clinical currency, it is clear that additional improvements in accuracy are needed, as few nomograms based on these parameters exceed a predictive accuracy of 80%. For example, the two most widely used preoperative nomograms that predict for PSA recurrence at 5 or 10 years after radical prostatectomy achieve accuracies of only 75% to 79%, despite more than 3,500 patients in the latter nomogram in the test and validation cohorts combined. Next-generation predictive tools incorporate additional biologic variables that may add more objective, tumor-specific information than can be achieved by clinical information and white-light histology alone. A recent study of a preoperative nomogram predicting likelihood of cure after radical prostatectomy demonstrated an improved accuracy of 88%, compared with 71% for clinical variables alone, by the addition of serum transforming growth factor 1 and interleukin-6-SR levels to the nomogram, validating a prior report. For breast cancer, the biologic approach has already yielded three commercially licensed products that rely on gene expression profiles (MammaPrint, Agendia BV, Amsterdam, the Netherlands; Oncotype DX, Genomic Health, Redwood City, CA; and H/I AvariaDX, Carlsbad, CA). In this issue of Journal of Clinical Oncology, Donovan et al take a different approach by adding cellular and biologic features to standard clinical and histologic parameters, a method called systems pathology. Using image analysis and quantitative immunofluorescence of hematoxylin and eosin–stained tissue microarrays and outcome data on 881 patients treated with radical prostatectomy, they identified 27 features reflecting color, texture, and the area of epithelial and stromal cells. Twelve of these were associated with a meaningful composite end point of clinical failure (defined as radiographic or pathologic evidence of metastasis, an increasing PSA in a castrate state, or death attributed to prostate cancer). Using support vector regression for censored data, a final model containing six variables (androgen receptor levels, prostatectomy Gleason grade, lymph node involvement, and three quantitative characteristics) predicting clinical failure within 5 years of surgery was generated. In the test set, the model had a concordance index of 0.92, sensitivity of 90%, and specificity of 91%, whereas in the validation set, the concordance index was 0.84, with a sensitivity of 84% and specificity of 85%. When compared with the JOURNAL OF CLINICAL ONCOLOGY E D I T O R I A L VOLUME 26 NUMBER 24 AUGUST 2