Estimating Treatment Effects from Irregular Time Series Observations with Hidden Confounders