Differences in cohort study data affect external validation of artificial intelligence models for predictive diagnostics of dementia - lessons for translation into clinical practice

Abstract

Artificial intelligence (AI) approaches pose a great opportunity for individualized, pre-symptomatic disease diagnosis which plays a key role in the context of personalized, predictive, and finally preventive medicine (PPPM). However, to translate PPPM into clinical practice, it is of utmost importance that AI-based models are carefully validated. The validation process comprises several steps, one of which is testing the model on patient-level data from an independent clinical cohort study. However, recruitment criteria can bias statistical analysis of cohort study data and impede model application beyond the training data. To evaluate whether and how data from independent clinical cohort studies differ from each other, this study systematically compares the datasets collected from two major dementia cohorts, namely, the Alzheimer's Disease Neuroimaging Initiative (ADNI) and AddNeuroMed. The presented comparison was conducted on individual feature level and revealed significant differences among both cohorts. Such systematic deviations can potentially hamper the generalizability of results which were based on a single cohort dataset. Despite identified differences, validation of a previously published, ADNI trained model for prediction of personalized dementia risk scores on 244 AddNeuroMed subjects was successful: External validation resulted in a high prediction performance of above 80% area under receiver operator characteristic curve up to 6 years before dementia diagnosis. Propensity score matching identified a subset of patients from AddNeuroMed, which showed significantly smaller demographic differences to ADNI. For these patients, an even higher prediction performance was achieved, which demonstrates the influence systematic differences between cohorts can have on validation results. In conclusion, this study exposes challenges in external validation of AI models on cohort study data and is one of the rare cases in the neurology field in which such external validation was performed. The presented model represents a proof of concept that reliable models for personalized predictive diagnostics are feasible, which, in turn, could lead to adequate disease prevention and hereby enable the PPPM paradigm in the dementia field.

Keywords: Alzheimer’s disease; Artificial intelligence; Bioinformatics; Cohort comparison; Cohort data; Data science; Dementia; Digital clinic; Disease modeling; Disease risk prediction; Health data; Individualized patient profiling; Interdisciplinary; Machine learning; Medical data; Model performance; Model validation; Multiprofessional; Neurodegeneration; Precision medicine; Predictive preventive personalized medicine (3 PM/PPPM); Propensity score matching; Risk modeling; Sampling bias; Survival analysis; Translational medicine.

Fig. 1

Caliper-based propensity score matching. (A) Procedure of caliper-based nearest neighbor propensity score matching as it was used in the comparison of ADNI and AddNeuroMed. The first step in the matching process is the calculation of a propensity score for each patient, followed by the matching of patients based on a caliper. The results are two cohorts consisting of patients similar with respect to the chosen matching features. Patients without match are discarded. (B) Caliper-based PSM as it was used for model validation. Only AddNeuroMed patients that found a match in ADNI were kept and used to validate the dementia risk model

Fig. 2

Distribution of propensity scores before and after PSM. (A) Scores for the full unmatched cohorts. (B) Scores for matched patients using a caliper of 1. (C) Scores for matched patients using a caliper of 0.5. ANM AddNeuroMed

Fig. 3

Performance of the dementia risk model on external validation and matched AddNeuroMed data calculated for 100 different matchings. (A) Harrell’s C-index of the model. The red line is indicates the model performance on the full unmatched AddNeuroMed cohort. (B) Area under the ROC over time (AUC(t)) showing the predictive performance over time before diagnosis. The standard error is plotted around the mean trajectory