You have several problems here. Firstly the signal is predominantly coming from your substrate. Your film is too thin to give an accurate measurement. An approach to solve that would be to reduce the voltage in order to concentrate the beam/specimen interaction within the film. However, to reduce it sufficiently would limit you to using Ti-L and O-K, where there is a significant overlap issue. One of the main assumptions of EDS is that the specimen should be homogeneous within the interaction volume - you do not have this. Secondly you have normalised data, that means you have not used standards to calibrate the system. I am guessing this is using virtual standards. This is acceptable for heavier elements, but when it comes to trying to measure light elements, such as O, then it becomes a problem, and a standardised calibration is necessary. However with the problem of the film thickness it becomes moot whether this would be worth it.

You may be better off trying XPS - I do not have experience with that, but someone else may be able to suggest its applicability.

Thanks a lot for your reply Ian.  You are right is pointing out the films being too thin (appx 50 nm) and we also didn't calibrate the system for this specific run.  So essentially I should not  estimate stoichiometry from this data.

See the paper: Stoichiometric determination of thin metal oxide films by S.Fazinić, I. Bogdanović a, E. Cereda, M. Jaksić and V. Valković ” available on Researchgate

Thanks Florian