With the increasing use of the electrocardiographic signal (ECG) as a diagnostic tool in cardiology, there exists a requirement for effective ECG compression techniques. The goal of any data compression system is to maximize compression while minimizing distortion. Orthogonal expansions is a tool widely used because of its compression capacity in recurrent signals. In this paper we analyze the effect of noise in orthogonal expansions of ECG signals. When the observed signal is embedded in additive noise, distortion measurements, such as the mean-square error, are not a monotonic decreasing function of the number of transform coefficients, due to the noise presence. We analyze and compare two different ways to estimate the transform coefficients: inner product and adaptive estimation with the LMS algorithm. For stationary signals, we demonstrate and quantify the superior performance obtained by the adaptive system when low values of the step-size are used μ<μlim. For non-stationary signals, we propose, based on experimental results, values of the LMS step-size μ depending on the noise characteristics and the signal-to-noise ratio. Theoretical results are contrasted with a simulation study with actual ECG signals from MIT-BIH Arrythmia database and three kinds of noise: simulated Gaussian white noise, and two records of physiological noise that essentially contains electrode motion artifacts and muscular activity.