Patient motion during cardiac SPECT imaging can cause diagnostic imaging artifacts. We have implemented a Neural Network (NN) approach to decompose monitored patient motion data, gathered during cardiac SPECT imaging, using the Polaris stereo-IR real-time motion-tracking system. Herein, we show the successful decomposition of Polaris motion data into rigid body motion (RBM) and respiratory motion (RM). The motivation for separating RM from RBM is that each is corrected using different methods. The NN requires the input of a RBM threshold sensitivity limit, as well as the median filter window width. A two step approach can be used in setting the median filter width. In the 1(st) NN run the median filter window width is initially set to a "fixed" width typical of the respiration period. This 1(st) NN run does an initial decomposition of the data into RM and RBM. The RM is then fed into an FFT algorithm to produce a respiratory period output file for use during a 2(nd) NN run, where the median filter width can "adapt" to the patient respiratory rate at each time point. Implementation of the NN was in the UNIX environment with Interactive Data Language (IDL). Decomposition of simulated "signals known exactly" RBM and RM resulted in average value errors less than 0.11 mm for RBM steps, and an overall root mean square error of only 0.3 mm for RM or RBM. Volunteer RBM and RM Polaris data were successfully decomposed by the NN with RBM steps resolved with an average difference of only 0.8 mm as compared to values displayed on the SPECT gantry console which are only to the nearest mm. A plot of the NN RM trace and the synchronized trace from a pneumatic bellows shows virtually identical characteristics. Anthropomorphic phantom RBM and RM were decomposed and used to correct motion in SPECT images during reconstruction. The motion corrected slices looked visually identical to slices acquired without motion, and comparison of slice count profiles further confirmed the correction.