This paper describes the prediction of disruptions based on diagnostic data in the high-β spherical torus NSTX (Ono et al 2000 Nucl. Fusion 40 557). The disruptive threshold values on many signals are examined. In some cases, raw diagnostic data can be used as a signal for disruption prediction. In others, the deviations of the plasma data from simple models provides the information used to determine the proximity to disruption. However, no single signal or calculation and associated threshold value can form the basis for disruption prediction in NSTX; thresholds that produce an acceptable false-positive rate have too large a missed or late-warning rate, while combinations that produce an acceptable rate of missed or late warnings have an unacceptable false-positive rate. To solve this problem, a novel means of combining multiple threshold tests has been developed. After being properly tuned, this algorithm can produce a false-positive rate of 2.8%, with a late + missed warning rate of 3.7% and thus a total failure rate of 6.5%, when applied to a database of ∼2000 disruptions during the IP flat top collected from three run campaigns. Furthermore, many of these false positives are triggered by near-disruptive magnetohydrodynamic (MHD) events that might indeed be disruptive in larger plasmas with more stored energy. However, the algorithm is less efficient at detecting the MHD event that prompts the disruption process.