Irregular time-series (ITS) are prevalent in the electronic health records (EHR) as the data is recorded in EHR system as per the clinical guidelines/requirements but not for research and also depends on the patient health status. ITS present challenges in training of machine learning algorithms, which are mostly built on assumption of coherent fixed dimensional feature space. In this paper, we propose a computationally efficient variant of the transformer based on the idea of cross-attention, called Perceiver, for time-series in healthcare. We further develop continuous patient state attention models, using the Perceiver and the transformer to deal with ITS in EHR. The continuous patient state models utilise neural ordinary differential equations to learn the patient health dynamics, i.e., patient health trajectory from the observed irregular time-steps, which enables them to sample any number of time-steps at any time. The performance of the proposed models is evaluated on in-hospital-mortality prediction task on Physionet-2012 challenge and MIMIC-III datasets. The Perceiver model significantly outperforms the baselines and reduces the computational complexity, as compared with the transformer model, without significant loss of performance. The carefully designed experiments to study irregularity in healthcare also show that the continuous patient state models outperform the baselines. The code is publicly released and verified at https://codeocean.com/capsule/4587224