Comparison Between External and Internal Imputation of Missing Values in Longitudinal Data for Alzheimer's Disease Diagnosis

Alzheimer's Disease (AD) presents a significant challenge for clinicians in both diagnosis and prognosis, with its effects rippling through society and healthcare systems. While advancements in treatment hold promise, early detection remains pivotal, demanding innovative technological solutions to enhance diagnostic and prognostic accuracy. Utilizing cognitive and neuroimaging data, researchers employ computational models, particularly artificial intelligence systems, to predict AD progression risk. The Alzheimer's Disease Neuroimaging Initiative (ADNI) provides a valuable dataset for such research; however, challenges persist, notably in handling missing data within longitudinal studies. This paper investigates the impact of imputation techniques on the performance of longitudinal deep learning methods for AD progression prediction. Four models-RNN, LSTM, DeepRNN, and ODE-RGRU-are evaluated with six imputation strategies, encompassing both external (five of them) overtures and an internal approach. External imputation methods of varying complexity are contrasted against the internal imputation capabilities of DeepRNN and ODE-RGRU. Our results highlight the efficacy of sophisticated methods like ODE-RGRU and DeepRNN, particularly when coupled with external imputation (ODE-RGRU with median imputation reached an mAUC value of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{0.9}\pm \mathbf{0.002}$</tex>, and DeepRNN using MissForest imputation reached an mAUC of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{0.91} \pm \mathbf{0.004}$</tex>). The findings underscore the importance of robust imputation techniques in enhancing the accuracy of AD progression prediction models, offering valuable insights for future research and clinical practice. These insights extend beyond AD research, particularly benefiting the burgeoning field of telemedicine, where missing data issues frequently occur.