Stochastic Models and the Early Detection of Disease

		   Marvin Zelen, Harvard School of Public Health 



Programs for the early detection of chronic diseases are proliferating. This is especially
true in cancer where there are early detection programs for breast, colon and prostate 
cancers.  The theory is that earlier detection of disease combined with beneficial therapy 
will lead to more cures or longer survival.  This talk will review developments in the 
stochastic modeling of the early detection of disease. The main issues relate to:  
evaluation of scientific evidence of benefit, the planning of early detection clinical 
trials and the planning of public health programs. It is possible that subjects found by 
an early detection program may be observed to live longer. However this may be an artifact 
due to lead and length time biases. The lead-time bias results from finding disease earlier,
but there is no benefit to the patient. The length bias arises because subjects diagnosed 
in such programs have more benign disease. Evaluation can only be done in the context of 
a randomized clinical trial. However the planning of such trials must take account features 
that do not arise in therapeutic trials; i.e. Number and spacing of exams and the optimal 
time for carrying out analyses.  Long-term follow up results in a loss of statistical power.
Public health programs are concerned with: populations which should be targeted, the age to 
begin such programs, schedules for high risk subjects and the number and spacing between 
early detection examinations. All of these problems rely on assumed stochastic models. 
Recent developments in this area will be discussed including: scheduling algorithms, 
innovative study designs for planning clinical trials, Group randomization and the 
estimation of examination sensitivity. The results presented here are joint work with: 
Ping Hu (National Cancer Institute), Sandra J. Lee (Dana-Farber Cancer Institute), Yu Shen 
(M. D. Anderson Hospital and Tumor Center).