For complex systems, the system failures exhibit an exponential nature, negating the benefit of traditional, time-based preventive maintenance. Under this circumstance, the condition-based predictive maintenance (CbPM) is more appropriate where the system condition can be monitored from the surface of running machinery, and maintenance is performed only when needed as the failure prognosis dictates. This paper presents a cost optimal prognostic criterion for CbPM using a multiple logistic function of risk variables to be monitored, which fluctuate randomly according to a certain probability distribution. A numerical example demonstrates the procedure and utility of the method.