This paper discusses the design and implementation of a comprehensive sensor based framework for medical activity recognition of elderly patients. Tri-axial accelerometers worn on the wrists monitor hand movements to recognize activities. A comparison of different methods of feature extraction from tri-axial accelerometer signals is performed. These are based on well known Time and Frequency domain analyses, Wavelet Transform and Hilbert-Huang Transform (HHT) using Empirical Mode Decomposition (EMD). The dataset of activities is collected from 30 subjects and comprises of the following activities: opening a pill box, popping a pill in the mouth, drinking water, using a syringe, falling down, sleeping, sugar test and BP test. Acitivites are classified using Naïve Bayes classifier, k-Nearest Neighbor algorithm and SVM algorithm. The activity recognition system is integrated with a GSM module that is used to automatically prompt appropriate alerts to a medical examiner in case of any abnormalities in the patient’s physical activities. This will have huge potential in reducing the cost of providing health care services to the elderly.