Artificial neural network coding of the child attachment interview using linguistic data

Assessing attachment in adolescents is important due to relations between insecurity and psychopathology. The child attachment interview (CAI) holds promise in this regard, but is time-consuming to code, which may render it inaccessible. The aim of this study was to develop computerized neural network models to predict attachment classifications on the CAI and to determine whether the models could achieve the CAI's benchmark qualification of 80% on reliability training cases. Four hundred and ninety interviews from inpatient adolescents served as model training and testing samples. The CAI's 30 standard reliability cases were treated as the independent holdout sample, in which the performance of the final models was evaluated against the 80% benchmark. Models demonstrated moderate accuracy and high correct classification rates, as compared to human coders. Performance was poorer when models were applied to the reliability training cases, but automated coding of the CAI holds promise for future development.