Text this: Temporal self-attention for risk prediction from electronic health records using non-stationary kernel approximation