Automatic prediction of cognitive assessment scores through analysis of speech is a challenging task not least due to the lack of available data; this is exacerbated by datasets often being accompanied by disparate cognitive scores as diagnostic practices vary across the world. The ADReSSo 2021 challenge aimed at supporting research in this area and defined a number of tasks including a regression task (predicting Mini-Mental State Examination (MMSE) scores). It saw the successful introduction of a number of BERT-based models including our winning classification approach that successfully applied data augmentation using ASR-generated hypotheses. In this paper, we port this approach to the regression task and further present an investigation into the effect of combining smaller datasets with disparate cognitive scores. In particular, we combine the ADReSSo data with our in-house IVA dataset, which is associated with a different type of cognitive assessment: the Addenbrooke's Cognitive Examination (ACE-III). We show improved performance by converting ACE-III to MMSE scores thus enabling us to combine the two datasets. By selecting good hyper-parameters, the RMSE reduces from 4.45 to 4.40 on the ADReSSo task. Likewise, using the ADReSSo dataset to boost the IVA regression model, decreases RMSE from 3.50 to 3.00.