Modelling the risk of conversion from mild cognitive impairment to Alzheimer’s disease using machine learning

Abstract

BACKGROUND:

Mild cognitive impairment (MCI) is a transitional state between normal aging and dementia, characterized by subtle cognitive decline. While not all individuals with MCI will progress to Alzheimer’s disease (AD), early identification of those at high risk is crucial for timely interventions. Machine learning algorithms offer a promising approach for analyzing complex datasets, including neuroimaging and cognitive assessments, to predict the likelihood of progression to AD. In this study, we evaluate 12 Machine Learning algorithms to predict the risk of progression from MCI to AD and identify the most important features contributing to this progression.

MATERIALS AND METHODS:

We selected 899 MCI subjects (596 stable, 303 progressive) from the AD Neuroimaging Initiative dataset. ML algorithms were used to predict the progression based on predictive features, including cognitive measures, demographic variables, and neuroimaging biomarkers.

RESULTS:

Linear Discriminant Analysis was identified as the top-performing model, achieving an accuracy of 73.43%, an area under the curve of 0.78, a recall of 69.53%, a precision of 59.38%, and an F1-score of 63.91%. These results demonstrate the model’s ability to effectively identify individuals at risk of progressing to AD. Hippocampal volume and the AD Assessment Scale-Cognitive were identified as the most influential predictors of conversion.

CONCLUSION:

The methodology can be generalized to other types of disease progression, such as transitions from healthy control to MCI or AD. Future studies could incorporate additional biomarkers and longitudinal data to enhance predictive accuracy and disease prevention.

Introduction

Alzheimer’s disease (AD) is a neurodegenerative disorder characterized by the deposition of neurofibrillary tangles and senile plaques, which leads to memory impairment and executive dysfunction. As the leading cause of dementia, AD is a significant contributor to mortality.[1,2]

The fifth edition of the Diagnostic and Statistical Manual of Mental Disorders (DSM-5)[3] brought about notable revisions in the categorization of neurocognitive disorders, particularly AD. Within this framework, AD is classified as a continuum under major and mild neurocognitive disorders, which enables a standardized diagnostic approach.[4] This taxonomy diverges from earlier, cause-centric systems, such as McKhann criteria[5,6] by emphasizing cognitive decline. The DSM-5 criteria for AD require decline from a prior functional level, with deficits in at least one cognitive domain,[4] and differentiate between mild and major neurocognitive disorders for earlier identification of lesser impairments.[7]

The comprehension of mild cognitive impairment (MCI) has evolved significantly, guided primarily by classification frameworks such as Petersen’s criteria[8] and the DSM-5 criteria. Both provide a unique approach to categorizing amnestic and non-amnestic MCI.

Petersen’s criteria categorize MCI into amnestic (aMCI) and non-amnestic (naMCI). While both subtypes may involve impairments in single or multiple cognitive domains, the amnestic variant is primarily characterized by memory decline. The DSM-5’s mild NCD offers a clinically useful framework for diagnosing the symptomatic pre-dementia phase of AD, emphasizing its pathological nature and functional impact. Combining medical and neuropsychological assessments with biological and neuroimaging biomarkers is pivotal for early detection and tracking disease progression.[9]

ML algorithms have demonstrated better performance than human-based evaluation metrics, underscoring their capacity to provide accurate predictions of AD.[10] AD can now be diagnosed by extracting a wide array of quantitative features from diverse medical imaging, such as electroencephalography, magnetoencephalography, positron emission tomography, and magnetic resonance imaging (MRI). The reviewed research predominantly utilized data from databases including Minimal Interval Resonance Imaging in AD, OASIS (Open Access Series of Imaging Research), and AD Neuroimaging Initiative (ADNI).[11,12]

In terms of neuropsychological testing, AD involves using standardized cognitive tests like the AD Assessment Scale-Cognitive (ADAS-Cog) and the Mini-Mental State Examination (MMSE).[13,14] Both are reliable for tracking disease progression. Neuropsychological assessment is crucial for identifying and characterizing the specific type of dementia linked to neurodegenerative conditions such as AD.[11]

The present study aims to (1) evaluate the risk of conversion from MCI to AD using only baseline data (i.e., first follow-up), and (2) investigate the contribution of cognitive test scores, demographic variables, and neuroimaging biomarkers to the probability of progression from MCI to AD by comparing the performance of 12 ML algorithms.

Materials and Methods

Study design and setting

Data used in the preparation of this article were obtained from the ADNI database (adni.loni.usc.edu) on June 2, 2025. Then, a sample of 899 individuals diagnosed with MCI across all ADNI phases (ADNI1, ADNI2/GO, ADNI3) was analyzed. All selected participants had a minimum of four years of follow-up data. The issue of missing data was addressed by excluding subjects with incomplete records rather than using imputation. After subjecting the first follow-up data to the optimal predictive model, we assessed the relative importance of its predictive features in determining the conversion risk from MCI to AD.

Study participants and sampling

This study initially enrolled 2430 subjects with at least four follow-up assessments. Only individuals with MCI who either progressed or remained stable throughout all follow-ups were selected. All participants in this study were independent of one another. However, if a patient has multiple missing values for a specific variable, it may affect other patients. Hence, we opted to exclude subjects with incomplete data rather than employing imputation methods. As a result, only subjects with complete datasets were enrolled in this study. Eight features were selected for analysis, including age, education, ADAS-Cog scores, MMSE scores, hippocampal volume (HC), whole brain volume (WBV), and entorhinal cortex volume (ECV). This approach ensures the integrity and reliability of the dataset used for subsequent analyses. Following this procedure, a sample size of 899 subjects was obtained, which is sufficient for all the ML employed in this study.

Because of the significant class imbalance in our dataset, with 303 progressive MCI (pMCI) subjects compared to 596 stable MCI (sMCI) subjects, the Synthetic Minority Oversampling Technique (SMOTE) was used to address this issue. This procedure generates synthetic samples for the minority class. This approach ensures a more balanced class during model training.[15]

The subject selection process is illustrated in the flowchart [Figure 1]. From the 2430 participants initially considered, 1551 were excluded due to missing data or because they were diagnosed with AD or CTRL at the first follow-up. As a result, a cohort of 899 individuals diagnosed with MCI was obtained, including 303 classified as pMCI and 596 as sMCI.

Figure 1.

Flow chart of the participant selection process

Data collection tool and technique

The ADNI study collects crucial data on disease progression. This includes tracking the time it takes for individuals with MCI to advance to probable AD, as well as the conversion of cognitively healthy individuals (CTRL) to either MCI or probable AD. Additionally, ADNI also gathers repeated longitudinal measurements (every 6 to 12 months) of patient-specific biomarkers, such as HCs and other neurodegeneration indicators.

Measures and variables

This study evaluated 12 ML algorithms to predict the progression from MCI to AD, incorporating selection based on eight features such as age, education, sex, HC, WBV, ECV, ADAS-Cog, and MMSE scores. Age, sex, and educational attainment, known to influence both AD risk and progression, were used in the analysis of MCI to AD conversion.

The top-performing algorithm was then utilized for predicting MCI to AD conversion. At each follow-up, the diagnosis was assessed according to the National Institute of Neurological and Communicative Disorders and Stroke and the Alzheimer’s Association criteria for probable AD.[16] The normality of each numerical variable was assessed using the Shapiro-Wilk test. Normally distributed variables were compared between groups using Student’s t-test, while non-normally distributed variables were analyzed with the Wilcoxon rank-sum test. The Chi-square test was used to assess differences in sex distribution between groups.

Neuropsychological assessment

For neurocognitive assessment, ADAS-Cog[13] and MMSE[14] are frequently employed to evaluate the transition from MCI to AD. The ADAS-Cog serves as an extensive evaluation tool that examines multiple cognitive areas, such as memory, language, attention, and orientation, thereby offering an in-depth understanding of cognitive capabilities. In contrast, the MMSE tool provides a rapid evaluation of overall cognitive health and is particularly adept at detecting deterioration in MCI cases. The two assessments combined improve the accuracy and reliability of predicting the progression to AD to MCI.

Neuroimaging biomarkers

Neuroimaging biomarkers, specifically HC, ECV, and WBA, were selected to predict the progression from MCI to AD. These biomarkers are highly correlated with the neuropathological changes characteristic of AD.[17] The brain regions involved are particularly vulnerable to AD-related neurodegeneration, and their structural changes, such as atrophy and volume reduction, can be detected early in the disease course using MRI. For our analysis, we only included neuroimaging biomarkers with no missing values.

Ethical considerations

This study utilized the publicly available ADNI dataset. The terms of use for the ADNI dataset are stated on their official website (https://adni.loni.usc.edu/terms-of-use/). All participants in the ADNI study provided written informed consent, approved by the institutional review board of each participating center. Participants were informed that their data would remain confidential, anonymized, and could be used in scientific publications.

Statistical analysis methods

ML algorithms have been widely applied across various medical specialties[18] and extract patterns that might not be immediately apparent to researchers.[19]

In this study, we employed the PyCaret library[20] to evaluate and compare 12 widely used ML algorithms, including Random Forest (RF), Logistic Regression (LR), Support Vector Machine (SVM), Linear Discriminant Analysis (LDA), Decision Tree (DT), Extra Trees (ET), Naïve Bayes (NB), Light Gradient Boosting Machine, K-Nearest Neighbors (KNN), Gradient Boosting (GB), AdaBoost (AB), and Quadratic Discriminant Analysis (QDA). All models were implemented using Python 3.11.10 within the Jupyter Notebook environment. Z-score normalization was applied to all numerical variables and computed as follows:

The dataset was randomly divided into a training set (80%, n = 719) and a test set (20%, n = 180). To mitigate the risk of overfitting, a 10-fold cross-validation strategy was employed.

The ML pipeline was implemented, encompassing data preprocessing, model setup, training, tuning, and performance evaluation. This workflow enabled systematic comparison of multiple algorithms under consistent conditions, ensuring reproducibility and robustness of predictive modelling. Model performance was assessed using common evaluation metrics, including accuracy, Area Under the Curve (AUC), recall, precision, and F1-score.[21] These metrics are computed based on the outcomes of the confusion matrix , where TP is the true positives, FP the false positives, TN the true negatives, and FN the false negatives. Each metric above was calculated as follows:

The F1-score provides a harmonic mean between precision and recall, offering a balanced measure especially in cases of imbalance. All the steps of the workflow of this study are represented in Figure 2.

Figure 2.

Machine learning workflow for Alzheimer’s disease progression

Results

Socio-demographic and neuropsychological characteristics of the two MCI groups (stable and progressive) enrolled in the study are presented in Table 1.

Table 1.

Comparative analysis of characteristics between stable and progressive mild cognitive impairment subjects. P values were obtained using Student’s t-test (or Wilcoxon rank-sum test) and the Chi-square test

Characteristics		Status=Stable MCI (n=596)		Status=Progressive MCI (n=303)		Statistical test		P
Sex
Female (%)		28.59		14.02		Chi-square test		<0.0001
Male (%)		37.71		19.69
Age (years)		71.84±7.50		73.15±7.13		t-test		0.0102
Education (years)		16.01±2.78		15.89±2.82		Wilcoxon test		0.5350
ADAS-Cog		8.56±3.73		12.57±4.56		Wilcoxon test		<0.0001
MMSE		27.94±1.81		27.03±1.86		t-test		<0.0001
Hippocampal		7151.63±1045.37		6312.99±1052.36		t-test		<0.0001
Whole brain		1051055±105568.2		1013598±113970.9		Wilcoxon test		<0.0001
Entorhinal cortex		3761.98±715.97		3271.79±785.97		t-test		<0.0001

ADAS-Cog=Alzheimer’s Disease Assessment Scale-Cognitive, MMSE=Mini-Mental State Examination

The proportion of females was significantly higher in the sMCI group (28.59%) compared to the pMCI group (14.02%) (P < 0.0001). Similarly, the proportion of males was significantly higher in the sMCI group (37.71%) compared to the pMCI group (19.69%) (P < 0.0001). At initial diagnosis, significant differences were observed between the stable and pMCI groups in terms of ADAS-Cog, MMSE, HC, WBV, and ECV (P < 0.0001). The difference was also significant in age (P = 0.01). However, no significant difference was found in years of education between the groups (P = 0.54).

Figure 3 presents the correlation among demographic variables, cognitive test scores, and neuroimaging biomarkers. ADAS-Cog shows a negative moderate correlation with HC (r = −0.4). Furthermore, the neuroimaging biomarkers are moderately to strongly intercorrelated (r = 0.5–0.6), reflecting common patterns of brain atrophy.

Figure 3.

Correlations between different features enrolled in the study

In Table 2, the LDA outperformed the other models evaluated. LDA achieved the highest accuracy of 73.43%, and the largest AUC of 0.78, alongside a recall of 69.53%, a precision of 59.38%, and an F1 score of 63.91%. These metrics represent the highest values, particularly in terms of accuracy and AUC, indicating that LDA is the most effective model among the 12 evaluated models.

Table 2.

Predictive performance of machine learning models for progression of mild cognitive impairment to Alzheimer’s disease

Model		Full Name		Accuracy		AUC		Recall		Prec.		F1
Lda		Linear Discriminant Analysis		0.7343		0.7775		0.6953		0.5938		0.6391
Lr		Logistic Regression		0.7301		0.7774		0.7077		0.5869		0.6402
Gbc		Gradient Boosting Classifier		0.7260		0.7392		0.5955		0.5980		0.5959
Nb		Naive Bayes		0.7219		0.7701		0.6873		0.5802		0.6262
Rf		Random Forest Classifier		0.7150		0.7394		0.5710		0.5838		0.5765
Qda		Quadratic Discriminant Analysis		0.7149		0.7529		0.6287		0.5753		0.5993
Et		Extra Trees Classifier		0.7094		0.7407		0.5418		0.5758		0.5564
lightgbm		Light Gradient Boosting Machine		0.7010		0.7348		0.5628		0.5573		0.5584
Svm		SVM - Linear Kernel		0.6997		0.7496		0.6293		0.5674		0.5853
Ada		Ada Boost Classifier		0.6968		0.7154		0.5953		0.5464		0.5680
Knn		K Neighbors Classifier		0.6648		0.7076		0.6533		0.5042		0.5676
Dt		Decision Tree Classifier		0.6328		0.6125		0.5502		0.4589		0.4990

AUC=Area under the curve

As shown in Figure 4, the ROC curves illustrate the classifier’s performance. For both Class 0 (Stable) and Class 1 (Progressive), the curves show an impressive AUC of 0.79, indicating the LDA classifier effectively distinguishes between these two classes individually.

Figure 4.

Receiver operating characteristic curves of the linear discriminant analysis classifier

Feature importance identification

Feature importance [Figure 5] measures the relative contribution of each variable, both numerical and categorical, in classifying the target within the training dataset.[22] The predictive significance of various factors in MCI progression was estimated using the LDA algorithms.[23]

Figure 5.

Feature importance plot of demographic, cognitive test, and biomarker characteristics of participants

The LDA model demonstrated the highest performance and stability when compared with all other models [Table 2]. ADAS-Cog and HC were the most significant predictors, while age and ECV also contributed substantially. These features are crucial due to their capacity to provide insights into relevant structural and functional alterations appropriate to the prediction task.[24] Conversely, sex, MMSE scores, education, and WBV ranked lower, indicating minimal impact on the model’s predictive performance. This trend highlights that model efficacy is primarily driven by anatomical brain measures and cognitive evaluation scores.

The confusion matrix illustrates the model’s performance in distinguishing between stable (Class 0) and progressive (Class 1) MCI. The model accurately classified 72% of sMCI cases and 67% of pMCI cases. These results demonstrate the model’s ability to differentiate between the two classes, despite some misclassification [Figure 6].

Figure 6.

Evaluation of the predictive performance of the linear discriminant analysis model on the test set

Discussion

To predict the progression from MCI to AD, we evaluated and compared the predictive performance of 12 widely used ML algorithms. The algorithms include LR, KNN, NB, DT, SVM, RF, QDA, AB, GB classifier, LDA, ET, and lightGBM. The LDA was the best selected model, with an AUC of 0.78 and an accuracy of 73.43%, highlighting the potential of ML in supporting early AD diagnosis. These results demonstrated the potential of ML in enhancing early AD diagnosis. Hence, the use of ML algorithms for disease prediction offers an appropriate alternative to traditional clinical assessments.

By including demographics (age, sex, and education), cognitive tests (ADAS-Cog and MMSE), and neuroimaging biomarkers (HC, WBV, and ECV) features, the selected model was able to identify subtle patterns that are often difficult to detect through human interpretation alone. Feature importance analysis [Figure 5] revealed that HC and ADAS-Cog scores have a strong effect on progression. These findings are consistent with prior literature.[25]

To evaluate classifier performance, the data were randomly divided into an 80% training set and a 20% testing set. The predictive models’ usage holds significant potential in stratifying MCI patients based on their risk of progression, enabling personalized monitoring and intervention strategies. Early identification of individuals at risk, particularly those transitioning from MCI to AD, can facilitate more targeted therapeutic interventions. This approach has the potential to improve patient outcomes by enabling timely interventions that may enhance early disease prevention and slow its progression.[26]

Regarding class imbalance, the SMOTE was applied, while Z-score normalization was used to standardize variations in data scale. These procedures are crucial for ensuring the generalizability of findings and the resilience of predictive models.

Previous studies describe aMCI patients who face a significant risk of progressing to AD, with conversion rates varying across various studies and diagnostic criteria. For instance, one investigation revealed that an annualized conversion rate of 28% is recorded within individuals evolving from aMCI to AD.[27] Another research identified three-year conversion rates of 6%, 53%, and 91% among low, medium, and high-risk groups, respectively, utilizing a point-based prediction tool.[28]

However, conversion rates for aMCI subtypes vary significantly. Patients with amnestic-plus MCI showed a notably elevated three-year conversion rate of 50%, compared to only 10% in those with pure aMCI.[29] Conversely, another study reported conversion rates of 48.7% for aMCI and 26.8% for naMCI after 30 months, indicating frequent progression to AD for both subtypes. Furthermore, structural MRI markers such as hippocampal and ECV have been consistently associated with progression from MCI to AD.[17] These findings are consistent with our results.

ADNI studies have significantly enhanced our understanding of the MCI to AD transition, especially for amnestic forms. Neuroimaging analyses revealed that MCI patients converting to AD within a year showed atrophic changes similar to those in established AD. Specifically, medial temporal lobe neurodegeneration, particularly reduced HC, was identified as the most reliable MRI predictor of conversion.[30]

Limitation and recommendation

A key strength of our study is using baseline follow-up data for conversion predictions. This approach simulates real-world clinical scenarios where early identification based on limited timepoints is needed. However, it also represents a limitation, as it excludes longitudinal measures that could enhance model accuracy.

Our Feature Importance analysis identified HC, ADAS-Cog scores, ECV, age, sex, education, WBV, and MMSE scores as significant risk factors for MCI progression, drawing from demographic, neuropsychological, and neuroimaging biomarkers. The ML model evaluates feature significance based on statistical patterns in the dataset rather than predetermined clinical assumptions. The application of predictive models shows substantial promise for patient stratification and early disease identification, enabling focused interventions and individualized care plans.

Early detection of cognitive decline can help with prompt treatment interventions. ML’s impact in contemporary neurology is shown by its incorporation into AD research. Another limitation is the absence of genetic data, such as APOE ε4 status, which was excluded due to missing values. Such information could provide insights into individual susceptibility and disease progression, enhancing model predictive performance.

While the LDA model outperformed other classifiers, further validation on other datasets is essential. Future directions should explore dynamic modelling approaches, such as joint modelling,[31] that can handle longitudinal data and non-linear interactions more effectively.

Conclusion

This study evaluated the effectiveness of ML algorithms in predicting the progression from MCI to AD, using neuroimaging biomarkers, demographic variables, and cognitive assessments. Among the 12 ML models tested, LDA showed the highest predictive performance, with an accuracy of 73.43% and an AUC of 0.78. Our findings highlight the relevance of HC, ADAS-Cog scores, and ECV as key predictors of disease progression. These biomarkers, when combined with ML techniques, hold promise for improving early AD diagnosis.

Further research should expand to more diverse cohorts, incorporate longitudinal data, and explore additional biomarkers to enhance predictive accuracy and disease prevention.

Conflicts of interest

There are no conflicts of interest.

Funding Statement

Nil.