Accelerometers are commonly used in clinical and epidemiological research for more detailed measures of physical activity and to target the limitations of self-report methods. Sensors are attached at the hip, wrist and thigh, and the acceleration data are processed and calibrated in different ways to determine activity intensity, body position and/or activity type. Simple linear modelling can be used to assess activity intensity from hip and thigh data, whilst more advanced machine-learning modelling is to prefer for the wrist. The thigh position is most optimal to assess body position and activity type using machine-learning modelling. Frequency filtering and measurement resolution needs to be considered for correct assessment of activity intensity. Simple physical activity measures and statistical methods are mostly used to investigate relationship with health, but do not take advantage of all information provided by accelerometers and do not consider all components of the physical activity behaviour and their interrelationships. More advanced statistical methods are suggested that analyse patterns of multiple measures of physical activity to demonstrate stronger and more specific relationships with health. However, evaluations of accelerometer methods show considerable measurement errors, especially at individual level, which interferes with their use in clinical research and practice. Therefore, better objective methods are needed with improved data processing and calibration techniques, exploring both simple linear and machine-learning alternatives. Development and implementation of accelerometer methods into clinical research and practice requires interdisciplinary collaboration to cover all aspects contributing to useful and accurate measures of physical activity behaviours related to health.