Paper Reading Journals #14

Basic Info

Item	Content
Title	Automated phenotyping of mild cognitive impairment and Alzheimer’s disease and related dementias using electronic health record
Journal	International Journal of Medical Informatics
Author	Ruoqi Wei, Stephanie S Buss, Marta Fernandes, et al.
Year Published	2025
Link	https://doi.org/10.1016/j.ijmedinf.2025.105917
Geography of Study	USA
Area of Study	ADRD, health system modeling, Electronic Health Record (EHR)

Description

(Methods & Models)

3,626 outpatients’ data were integrated into the cohort. The cohort was composed by three types of data: (1) unstructured diagnosed data; (2) ICD codes; (3) medication prescriptions data related to MCI/ADRD. 4 different machine learning models were adopted to perform the classification task. These four ML models are: (1) logistic regression (2) random forest (3) naive Bayes (4) multi-layer perceptron (MLP). 5-fold stratified nested cross-validation and other general ML metrics (1. accuracy 2. specificity 3. F1-score 3. Recall 4. Precision 5. AUROC 6. AUPRC) were presented as validation. In result, the logistic regression have the best performance on MGH+BIDMC test set among all machine learning models. it achieve a 0.91 accuracy, 0.94 specificity, 0.92 precision, and 0.97 AUROC score. The author also reported top 15 features based on coefficient values. Top 5 are: “dementia”, “donepezil”, “cognitive impairment”, “alzheimers”, “alzheimer”.

Data Source

The retrospective cohort (n = 3,626) was collected from two local health systems: (1) Massachusetts General Hospital (MGH) (2) Beth Isreal Deaconess Medical Center (BIDMC). Data were collected for visits during Janurary 3, 2012 and November 3, 2017. Only patients aged 50 years or older were admit into cohort to reflect eh age-associated nature of MCI and ADRD.

Main Contribution

(Innovative method or Application)

Innovation Type: Application Innovation The main contribution of

Favorite Thing