If I understand, you want to make an adaptive threshold to T amplitude. This may work in normal sinus rhythm but probably will not give good results in noisy or pathological ECGs. Perhaps you should take the first derivative of the signal and then try to apply an average threshold to cancel T wave peaks. Try both and compare.

Here is one more approach in a paper published by Prof R B Ghongade of VIIT Pune India : Abstract of his paper published is available

A NOVEL QRS DETECTION METHOD

Rajesh Ghongade, V.I.I.T, Pune, INDIA

Dr. Ashok Ghatol, Dr. B.A.T.U., Lonere, INDIA

Abstract: The most important step for automated human ECG diagnosis is the reliable detection and extraction of QRS complexes. There are several issues which aggravate the problem of reliable QRS detection and those are; patient to patient variation in signal amplitudes, power-line noise infiltration, baseline or DC shift and muscle activity. Various algorithms are suggested and experimented with in the literature, but still a universal scheme for QRS detection for all types of ECG signals is non-existent. This paper presents a completely novel approach for QRS detection and extraction using energy-gradient technique. The technique was evaluated with all the 48 records from MIT-BIH Arrhythmia database. The method is found to be highly effective and robust and exhibits average detection accuracy of 98.94%, average sensitivity of 0.9894, and average positive predictivity of 0.987 and the average detection error rate as 2.26%. The proposed method is found to be most suitable for off-line ECG QRS detection.

you know the ECG signal is stochastic random signal therefore according to your application you can decide the suitable features which is suitable for your analysis if you deal with amplitude you can make normalized with respect to maximum value and used the threshed according to slope of tangent function. the a lot of application used the threshold more than 60% .