Four-level pulse amplitude modulation (PAM4) signals transmission in short-haul intensity modulation-direct detection datacenters connections supported by homogeneous weaklycoupled multicore fibers is seen as a promising technology to meet the future challenge of providing enough bandwidth and achieve high data capacity in datacenter links. However, in multicore fibers, inter-core crosstalk (ICXT) limits significantly the performance of such short-reach connections by causing large bit error rate (BER) fluctuations. In this work, a convolutional neural network (CNN) is proposed for eye-pattern analysis and BER prediction in PAM4 inter-datacenter optical connections impaired by ICXT, with the aim of optical performance monitoring. The performance of the CNN is assessed by estimation of the root mean square error (RMSE) using a synthetic dataset created with Monte Carlo simulation. Considering PAM4 interdatacenter connections with one interfering core and for different skewsymbol rate products, extinction ratios and crosstalk levels, the obtained results show that the implemented CNN is able to predict the BER without surpassing a RMSE limit of 0.1.