An aid diagnostic platform to detect the transition of mild cognitive impairment (MCI) to Alzheimer's disease (AD) based on 48,116 AD and MCI patients

doi:10.21203/rs.3.rs-4108664/v1

Download PDF

Article

An aid diagnostic platform to detect the transition of mild cognitive impairment (MCI) to Alzheimer's disease (AD) based on 48,116 AD and MCI patients

https://doi.org/10.21203/rs.3.rs-4108664/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this older preprint version

Read the latest preprint version →

Alzheimer's disease (AD) is an incurable, progressive neurodegenerative disorder, necessitating early diagnosis and intervention. Mild cognitive impairment (MCI) often precedes AD, but not all cases progress to AD, emphasizing the need for predictive biomarkers. We analyzed routine blood test data from 43,981 AD patients and 4,537 MCI subjects in Hong Kong hospitals (2000–2019). Among 31 shared biomarkers, five blood biomarkers (Hemoglobin, Hematocrit, Red blood cell related to oxygen carrying capacity, Neutrophils, and White blood cell related to immunity) significantly differentiated MCI from AD. Subjects were divided into four groups (Female 65 ~ 74, Male 65 ~ 74, Female 75 ~ 89, Male 75 ~ 89) to minimize gender and age bias. Models utilizing the five biomarkers along with machine learning yielded the highest accuracy in the Female 65 ~ 74 group (AUC of 0.76 on an independent test set). The other three models were trained with other biomarkers besides these 5 to optimize predictions, capturing models with AUC close to 0.70. We then constructed a platform predicting the risk of MCI converting to AD (MAP, http://lab.malab.cn/~lijing/MAP.html) to help physicians and MCI subjects with early diagnosis and prevention of AD. In conclusion, this study demonstrates the potential for accurate prediction of MCI to AD conversion using routine blood test data and machine learning, offering an economical and practical approach for early AD screening in MCI individuals.

Health sciences/Diseases/Neurological disorders/Dementia/Alzheimer's disease

Biological sciences/Computational biology and bioinformatics/Computational models

Health sciences/Biomarkers/Predictive markers

Alzheimer's disease (AD) is the most common cause of dementia in older people and is characterized by a gradual decline in memory and cognition. With an increasing global aging population, AD is becoming a significant health concern worldwide^1,2. By 2050, AD is expected to affect over 100 million people, with approximately 4.6 million newly diagnosed cases per year³. According to the Alzheimer's Association report in 2020, the total cost of each AD patient is about US$46,786 annually, and the global cost of the disease is estimated to exceed US$2 trillion by 2030⁴. Unfortunately, there are no effective drugs or interventions that can stop or reverse disease progression. This is partly because treatments are usually administered after the onset of dementia symptoms when AD pathology has already advanced. Therefore, early diagnosis and treatment of AD are essential for successful preventive and therapeutic strategies.

Mild cognitive impairment (MCI) has attracted clinical and research interest in recent years with the hypothesis that it represents a “transitional zone” between normal cognition and dementia⁵. Subjects with MCI show objective cognitive impairment but normal capabilities for daily activities. MCI subjects are at increased risk of developing AD, with an annual conversion rate of 10–30%, compared with 1–2% in unaffected individuals⁶. It is believed that administering intervention during the MCI stage will increase the probability of therapeutic success in AD. However, not all MCI subjects progress to AD, some cases even return to normal aging⁷. Hence, it is critical to identify those presymptomatic individuals with MCI who will eventually convert to Alzheimer’s, to yield substantial therapeutic and health-economic benefits.

Like AD diagnosis, current diagnostic methods of MCI are mainly based on neuropsychological tests and neuroimaging techniques, assisted with biomarkers (such as Aβ, tau, and phosphorylated tau) in cerebrospinal fluid (CSF) and blood. Many prediction models for MCI to AD conversion have been developed using the results of these methods^8–12. Nevertheless, there are some limitations to those techniques. Firstly, the clinical tests show low specificities and neuroimaging methods are costly. Despite the important diagnostic utility of CSF biomarkers, the detection requires an invasive lumbar puncture. Additionally, the measurement of blood Aβ peptides and tau is challenging as their concentrations are extremely low in the blood, and their applications in large-scale diagnostic and prognostic trials are limited due to the necessity of sensitive and specific analytical approaches.

Previous studies have demonstrated the potential of using routine peripheral blood parameters to predict AD risk as many parameters are altered in AD patients and some factors are associated with cognitive decline^13–19. Considering the easy accessibility and cost-effectiveness of routine blood tests, routine blood analytes could complement neuropsychological tests as screening tools before conducting neuroimaging tests and detecting CSF/blood biomarkers to predict future cognitive decline in MCI subjects. However, there are very few that explore the potential of utilizing routine laboratory analytes to predict the risk of conversion from MCI to AD.

Machine learning has been widely used in various fields, including auxiliary medical diagnosis^20–22. With the help of mathematical algorithms, machine learning can generate a function to predict the disease state by using patients’ laboratory test results. Recently, several efforts have been made to create AD risk prediction models using machine-learning methods^23–25, which can attain a more significant and accurate result than traditional methods.

This paper aims to establish an AD prediction model in MCI subjects through routine blood test data and machine-learning methods. With a large population of 43,981 patients with AD and 4,537 subjects with MCI and 210 original features gathered from routine blood tests accumulated from 2000 to 2019 in Hong Kong, we were able to generate a prediction model for AD conversion by using variable machine-learning algorithms. Additionally, analyses were further grouped by gender (male and female) and age (65 ~ 74 years old, 75 ~ 89 years old) to obtain a more accurate prediction with an AUC reaching 0.70. Our findings highlight the potential of using routine laboratory test data and sophisticated machine-learning methods for accurate AD risk modeling in MCI subjects and provide a useful approach for the early screening of AD.

Individual biomarker analysis

Our data was obtained from HADCL (Healthcare Analytics and Data Center). We identified 43,981 patients with AD from a population of 12 million individuals spanning the years 2000 to 2019. Within the dataset, comprising an initial set of 210 routine blood biomarkers, we conducted an in-depth analysis focusing on 31 specific biomarkers. For a comprehensive understanding of our selection process, please refer to the Methods section. Student’s t-test²⁶ was used to evaluate the differences in the 31 routine blood biomarkers between MCI and AD among different groups (using data that does not contain missing values). A significance threshold of p < 0.05 was set for the t-test results. Out of the 31 biomarkers subjected to t-test analysis, 24 exhibited a significant difference between MCI and AD patients, as indicated by p-values less than 0.05. These significant findings are presented in Table 1. The remaining 7 biomarkers, with p-values exceeding 0.05, did not show statistically significant differences and were not included in this table. Notably, hemoglobin (blood), HCT (Haematocrit), and RBC (Red Blood Cell) were significantly reduced, while neutrophils (absolute) and WBC (White Blood Cell) were increased in AD patients among all 4 groups.

Model development

As stated the data process in Method section, we finally excluded 10,344 patients, leaving total of 7159 patients with 777 patients in Female 65 ~ 74 (MCI, n = 131; AD, n = 646), 642 patients in Male 65 ~ 74 (MCI, n = 147; AD, n = 495), 3,821 patients in Female 75 ~ 89 (MCI, n = 249; AD, n = 3,572), and 1,919 patients in Male 75 ~ 89 (MCI, n = 222; AD, n = 1,697).

However, the number of AD patients far exceeds that of MCI patients in our data. Given that the imbalance in classification may result in high false positive rates and low generalization when modeling^27,28, the AD patients were randomly reduced to the same number as MCI subjects, to minimize the bias. The final number of each group after balancing was 262 patients in Female 65 ~ 74 (MCI, n = 131; AD, n = 131), 294 patients in Male 65 ~ 74 (MCI, n = 147; AD, n = 147), 498 patients in Female 75 ~ 89 (MCI, n = 249; AD, n = 249), and 444 patients in Male 75 ~ 89 (MCI, n = 222; AD, n = 222). For more details, please refer to Fig. 1.

To train more stable models, ensemble learning²⁹ was adopted in this work. Ensemble learning combines multiple weakly supervised models to create more comprehensive models. The algorithm assigns higher weights to the more important weak classifiers, and if a weak classifier predicts a patient wrong, the other weak classifiers can correct it. To establish the best classification algorithm, 85% of each data set (Female 65 ~ 74, Male 65 ~ 74, Female 75 ~ 89, and Male 75 ~ 89) was selected randomly as the training set, and the remaining 15% was used for the independent test set. Details for final data distribution are shown in Fig. 1. The 5-fold cross-validation (CV)³⁰ was used to train the weak classifier models, and extensive screening and grid search were applied to determine optimal algorithms and the best-performing weights, respectively.

According to the t-test results, there were 5 significant biomarkers in the 4 groups, including Hemoglobin (blood), HCT, Neutrophil (absolute), RBC, and WBC. Results from models trained with these 5 biomarkers showed that only the model of Female 65 ~ 74 had an AUC greater than 0.65. In Female 65 ~ 74, we combined the above 5 biomarkers with Logistic Regression (LR)³¹, Gaussian Naive Bayes (GNB)³², and Random Forest (RF)³³ to generate basic classifiers that were then applied to construct the Female 65 ~ 74 model. These basic classifiers were used for ensemble learning with a weight set of LR: GNB: RF = 1: 3: 1 (Fig. 2). In the other 3 groups, the performances of models using feature matrix formed by the 5 significant biomarkers could not achieve a relatively high accuracy (AUCs were lower than 0.65). We thus used forward feature selection (FFS)³⁴ to add biomarkers from Table 1 one by one to optimize the models. We finally found an optimal model for Female 75 ~ 89, which was generated by basic classifies integrated 19 biomarkers (analytes related to liver function (Alanine Aminotransferase (ALT), Alkaline Phosphatase (ALP), Bilirubin), minerals and proteins (Potassium, Calcium, Phosphate, Protein, Albumin), and blood cells (Basophil (absolute), Neutrophil (absolute), Lymphocyte (absolute), Basophil (%), Neutrophil (%), Lymphocyte (%), WBC, RBC, Mean corpuscular hemoglobin concentration (MCHC), Hematocrit (HCT, Hemoglobin)) with LR, Gaussian Naive Bayes (GNB)³⁵, and Extra Trees (ET)³⁶ algorithms. Similarly, ensemble learning was used to build the Female 75 ~ 89 model with a weight set of LR: GB: ET = 1: 3: 4 (Fig. 2).

Because the optimization by FFS did not work well for male models, we changed the strategy by using all 31 biomarkers to train the model. We ultimately constructed a relatively good prediction model for Male 65 ~ 74 through 31 biomarkers and Linear Discriminant Analysis (LDA)³⁷, Support Vector Machine (SVM)³⁸, and ET algorithm. Ensemble learning was used to construct the Male 65 ~ 74 model with a basic classifier weight set of LDA: SVM: ET = 1: 2: 4 (Fig. 2). As for Male 75 ~ 89, we first performed Maximum-Relevance-Maximum-Distance (MRMD)³⁹ feature selection to screen out 23 biomarkers (ALP, variables associated with renal function (Creatinine, Urea), minerals and proteins (Potassium, Sodium, Calcium, Phosphate, Albumin, Protein), and blood cells (Basophil (absolute), Neutrophil (absolute), Eosinophil (absolute), Lymphocyte (absolute), Monocyte (absolute), Basophil (%), Monocyte (%), WBC, Platelet, Mean corpuscular volume (MCV), Mean corpuscular hemoglobin (MCH), MCHC, Hemoglobin, Red blood cell distribution width (RDW) (%)), and then combined them with RF, AdaBoost Classifier (ADA), and ET algorithms. These basic classifiers were used for ensemble learning to generate the Male 75 ~ 89 model with a weight set of RF: ADA: ET = 1: 2: 3 (Fig. 2).

Routine blood biomarker-driven models and effective predictive performance

The prediction accuracy and AUC values of each model on the CV are shown in Table 2 and Fig. 3. Specifically, applying the ensemble methods combined with LR³¹, GNB³², and RF³³, the 5-biomarkers for the Female 65 ~ 74 model yielded an accuracy (ACC) of 0.63 and an AUC of 0.70, with sensitivity (SN) at 0.61 and specificity (SP) at 0.65. For the Female 75 ~ 89 model, the 19-biomarkers integrated with LR, GB³⁵, and ET³⁶ achieved similar performance, resulting in an ACC = 0.63, AUC = 0.67, SN = 0.69, and SP = 0.58. Likewise, the 31-biomarkers constructed with LDA³⁷, SVM³⁸, and ET for the Male 65 ~ 74 model provided an ACC at 0.63, AUC at 0.66, SN at 0.62, and SP at 0.64, and the 23-biomarkers generated with RF, ADA, and ET for the Male 75 ~ 89 model obtained an ACC = 0.66, AUC = 0.68, SN = 0.66, and SP = 0.65.

Table 2

The model performance on cross-validation test sets
		Cross-validation test sets
		SN	SP	ACC	F1	AUC
ML model	F 65 ~ 74	0.61±0.01	0.65±0.02	0.63±0.00	0.62±0.00	0.70±0.01
	TabNet	0.50±0.34	0.70±0.28	0.60±0.08	0.48±0.26	0.74±0.08
ML model	M 65 ~ 74	0.62±0.02	0.64±0.02	0.63±0.02	0.62±0.02	0.66±0.00
	TabNet	0.65±0.31	0.58±0.36	0.61±0.08	0.59±0.15	0.72±0.06
ML model	F 75 ~ 89	0.69±0.10	0.58±0.08	0.63±0.18	0.66±0.24	0.67±0.22
	TabNet	0.35±0.40	0.74±0.36	0.55±0.07	0.31±0.33	0.63±0.10
ML model	M 75 ~ 89	0.66±0.04	0.65±0.04	0.66±0.00	0.66±0.02	0.68±0.01
	TabNet	0.39±0.32	0.81±0.21	0.60±0.07	0.42±0.30	0.67±0.10
Abbreviations: SN, sensitivity; SP, specificity; ACC, accuracy; F1, F-score; AUC, area under the receiver operating characteristic curve; F 65 ~ 74, Female 65 ~ 74 years old; M 65 ~ 74, Male 65 ~ 74 years old; F 75 ~ 89, Female 75 ~ 89 years old; M 75 ~ 89, Male 75 ~ 89 years old.

We next evaluated the model performance on independent test sets. The prediction results were summarized in Table 3 and Fig. 3, and the detailed ROC curves were depicted in Fig. 4. Consistent with the results on CV test sets, the performances of each model on independent test sets were relatively stable, yielding an average ACC of 0.6425, AUC of 0.685, SN of 0.6275, and SP of 0.6575. The consistency of performance on CV test sets and independent test sets suggests that our models are resistant to noise, outliers, or slight changes in the data set and thus can provide a reliable prediction for unknown samples. Our model demonstrates robust consistency, withstanding variations such as noise and outliers, affirming reliability in predicting the progression from MCI to AD. This stability is crucial for application in clinical settings, where it can aid in the early detection of high-risk individuals.

Table 3

The model performance on independent test sets
		independent test sets
		SN	SP	ACC	F1	AUC
ML model	F 65 ~ 74	0.70±0.04	0.66±0.02	0.68±0.02	0.69±0.02	0.76±0.02
	TabNet	0.43±0.34	0.65±0.31	0.54±0.09	0.41±0.26	0.59±0.11
ML model	M 65 ~ 74	0.59±0.04	0.65±0.07	0.62±0.05	0.61±0.05	0.65±0.07
	TabNet	0.54±0.35	0.47±0.33	0.51±0.05	0.47±0.21	0.49±0.09
ML model	F 75 ~ 89	0.61±0.14	0.68±0.04	0.64±0.17	0.63±0.20	0.66±0.22
	TabNet	0.32±0.38	0.72±0.36	0.52±0.05	0.29±0.30	0.57±0.07
ML model	M 75 ~ 89	0.61±0.00	0.64±0.00	0.63±0.00	0.62±0.00	0.67±0.00
	TabNet	0.28±0.24	0.80±0.17	0.54±0.06	0.32±0.24	0.57±0.09
Abbreviations: SN, sensitivity; SP, specificity; ACC, accuracy; F1, F-score; AUC, area under the receiver operating characteristic curve; F 65 ~ 74, Female 65 ~ 74 years old; M 65 ~ 74, Male 65 ~ 74 years old; F 75 ~ 89, Female 75 ~ 89 years old; M 75 ~ 89, Male 75 ~ 89 years old.

Impact of age and gender on prediction performance

In MAP, we stratified patients into 4 distinct age and gender groups and constructed models trained with 5-, 31-, 19-, and 23-biomarkers corresponding to Female 65 ~ 74, Male 65 ~ 74, Female 75 ~ 89, and Male 75 ~ 89, respectively. To demonstrate the influence of age and gender stratification on the prediction from MCI to AD, we performed an additional comparative experiment using data from all individuals aged 65 ~ 89 years, irrespective of gender. In MAP, the 5-, 19- and 23-biomarkers used for training the Female 65 ~ 74, Female 75 ~ 89, and Male 75 ~ 89 models were derived from the initial 31 biomarkers according to a secondary selection process after age and gender stratification. For the non-stratified data, the 5- and 19-biomarker data were selected directly from all MCI and AD patients to increase the sample size for analysis, while the 23-biomarker data was strictly adhered to a two-step screening method, where they were chosen through feature selection using MRMD from the initial set of 31 biomarkers (Fig. 5). As described above, we also excluded the patients with missing values, balanced the data between AD and MCI groups, and divided the data into training and independent test sets for model training (Fig. 5).

Results of the non-stratified data showed a comparatively less optimal model performance when employing the same machine-learning methodology as MAP. As shown in Table 4, without age and gender stratification, the models exhibited higher SN but lower SP, leading to a bias in identifying MCI subjects who did not convert to AD. This phenomenon may be attributed to the model excessively focusing on the biomarkers of AD patients while neglecting those of MCI subjects. Consistent with the cross-validation findings, the independent test results also revealed that models without categorizing patients into different age and gender groups yielded inferior performance compared to models grouped by age and gender. Specifically, the former provided an average ACC of 0.56, an AUC of 0.605, a SN of 0.7175, and a SP of 0.4075 (Table 4). Our findings reveal the crucial role of age and gender stratification in improving the accuracy of identifying individuals at risk of progressing from MCI to AD.

Table 4

The model performance on data without age and gender stratification
Cross-validation test sets
data	model	SN	SP	ACC	F1	AUC
features	5 biomarkers
F & M 65 ~ 89	ML model	0.72±0.01	0.39±0.01	0.56±0.01	0.62±0.01	0.58±0.00
F & M 65 ~ 89	TabNet	0.50±0.18	0.64±0.13	0.57±0.05	0.52±0.13	0.61±0.05
features	19 biomarkers
F & M 65 ~ 89	ML model	0.73±0.01	0.73±0.01	0.73±0.01	0.73±0.01	0.60±0.00
F & M 65 ~ 89	TabNet	0.69±0.31	0.69±0.31	0.69±0.31	0.69±0.31	0.61±0.04
features	23 biomarkers
F & M 65 ~ 89	ML model	0.75±0.01	0.41±0.02	0.58±0.01	0.64±0.01	0.62±0.01
F & M 65 ~ 89	TabNet	0.49±0.39	0.54±0.43	0.52±0.03	0.44±0.19	0.50±0.06
features	31 biomarkers
F & M 65 ~ 89	ML model	0.74±0.01	0.42±0.01	0.58±0.01	0.64±0.01	0.63±0.01
F & M 65 ~ 89	TabNet	0.34±0.20	0.76±0.18	0.55±0.03	0.41±0.16	0.62±0.04
independent test sets
features	5 biomarkers
F & M 65 ~ 89	ML model	0.71±0.01	0.38±0.01	0.54±0.01	0.61±0.02	0.57±0.01
F & M 65 ~ 89	TabNet	0.48±0.18	0.58±0.15	0.53±0.02	0.49±0.11	0.55±0.01
features	19 biomarkers
F & M 65 ~ 89	ML model	0.73±0.01	0.43±0.01	0.58±0.01	0.63±0.02	0.62±0.02
F & M 65 ~ 89	TabNet	0.70±0.32	0.37±0.32	0.54±0.03	0.56±0.06	0.55±0.03
features	23 biomarkers
F & M 65 ~ 89	ML model	0.71±0.02	0.42±0.01	0.56±0.01	0.62±0.01	0.62±0.02
F & M 65 ~ 89	TabNet	0.55±0.37	0.56±0.41	0.56±0.04	0.50±0.17	0.59±0.02
features	31 biomarkers
F & M 65 ~ 89	ML model	0.72±0.01	0.40±0.01	0.56±0.01	0.62±0.01	0.61±0.01
F & M 65 ~ 89	TabNet	0.37±0.09	0.71±0.08	0.54±0.01	0.42±0.05	0.56±0.12
Abbreviations: F & M 65 ~ 89, Female and Male 65 ~ 89 years old.

Comparative analysis of MAP and TabNet models for predicting the risk of MCI to AD

To compare the performance of MAP with models trained with deep-learning methods, the 5-/19-/23-/31-biomarker sets were trained with TabNet⁴⁰, a deep-learning method that is widely used to train interpretable models, to construct TabNet models with or without age-and gender-grouping. Results revealed poor performances of TabNet models on CV test sets and independent test sets. When grouped by age and gender, TabNet models had similar AUCs as MAP on CV test sets, but their ACCs were lower and the differences between SNs and SPs were larger than MAP (Table 2, and Fig. 3). Regarding independent test sets, TabNet models obtained lower ACCs, AUCs, and SNs, but larger differences between SNs and SPs than MAP (Table 3, and Fig. 3, 4). When the data was not stratified by age and gender, TabNet models displayed slightly inferior performances as MAP on CV test sets and independent test sets (Table 4).

Participants

We investigated the routine laboratory data of registered patients with AD or MCI from the Hong Kong Hospital Authority Data Collaboration Laboratory (HADCL). Officially launched in December 2019, HADCL aims to provide health data to facilitate research and help improve clinical and healthcare services. AD and MCI diagnosis results were extracted from the main diagnosis field of the outpatient or inpatient records from 2000 to 2019 using the International Statistical Classification of Diseases and Related Health Problems (ICD-10)⁴¹ diagnostic codes G30.0, G30.1, G30.8, G30.9, and G31.84, assisted with proposed Mental and Behavioral Disorder (MBD) diagnostic definitions for ICD10-2010. A total of 48,116 registered patients, who were aged 65 years or older at diagnosis with AD or MCI, were identified in the platform data. These subjects were categorized into three groups: AD patients (43,579; Female 30,041; Male 13,538), who were diagnosed with AD at the first diagnosis; MCI subjects (4,135; Female 2,157; Male 1,978), those who were diagnosed with MCI and did not convert to AD before 2019; and MCI-to-AD patients (402; Female 237; Male 165), who were diagnosed with MCI at the initial state and then converted to AD at the follow-up time. The collected common laboratory data for each patient included chemical pathology analytes, hematology analytes, and immunology analytes, resulting in 210 analytes in total.

Data processing

Although there are 402 samples converted from MCI to AD, only 1 male subject and 1 female subject have at least three years of follow-up records. In addition, the exact timing of an MCI subject converting to AD is obscure. We thus removed the MCI-to-AD group due to the limited available data.

Considering the common laboratory data usually contain missing values, we used the analytes which are shared by more than 90% of subjects as the quality filter of samples. Those biomarkers include analytes related to liver function (Alanine Aminotransferase (ALT), Alkaline Phosphatase (ALP), Bilirubin), variables associated with renal function (Creatinine, Urea), minerals and proteins (Potassium, Sodium, Calcium, Phosphate, Albumin, Globulin, Protein), and blood cells (Eosinophil (absolute), Basophil (absolute), Neutrophil (absolute), Lymphocyte (absolute), Monocyte (absolute), Eosinophil (%), Basophil (%), Neutrophil (%), Lymphocyte (%), Monocyte (%), White blood cells (WBC), Platelet, Red blood cells (RBC), Mean corpuscular volume (MCV), Mean corpuscular hemoglobin (MCH), Mean corpuscular hemoglobin concentration (MCHC), Hematocrit (HCT), Hemoglobin, Red blood cell distribution width (RDW)). During this process, we each time employed one biomarker to screen out patients who have the value of this biomarker. After 31 rounds of screening, we collected all these screened-out patients for following analysis. Finally, a total of 30,211 samples were removed, resulting in 15,900 AD patients and 1,603 MCI subjects.

Gender and age are important risk factors for AD. To exclude their influence on the prediction results, the patients were further categorized into four groups: female subjects between the ages of 65 and 74 years old (Female 65 ~ 74; MCI, n = 262; AD, n = 1,622), male subjects between the ages of 65 and 74 years old (Male 65 ~ 74; MCI, n = 307; AD, n = 1,331), female subjects between the ages of 75 and 89 years old (Female 75 ~ 89; MCI, n = 517; AD, n = 8,760), and male subjects between the ages of 75 and 89 years old (Male 75 ~ 89; MCI, n = 517; AD, n = 4,187). The preliminary performance of the prediction models generated with samples containing less than 2 (including 2) missing values of the 31 biomarkers exhibited very poor stability (For more details, please refer to supplementary data), we thus removed these samples and only kept those without any missing values. To ensure the reliability of the outcomes, we used the Interquartile Range (IQR)⁴² to remove the outliers in the samples, which is widely adopted to find outliers among data by dividing the data set into quartiles and determining the IQR using the distance between the quartile 3 and quartile 1. Samples with moderate outliers (K = 1.5) were removed. Finally, we excluded 10,344 patients, leaving total of 7159 patients with 777 patients in Female 65 ~ 74 (MCI, n = 131; AD, n = 646), 642 patients in Male 65 ~ 74 (MCI, n = 147; AD, n = 495), 3,821 patients in Female 75 ~ 89 (MCI, n = 249; AD, n = 3,572), and 1,919 patients in Male 75 ~ 89 (MCI, n = 222; AD, n = 1,697).

Machine learning algorithm

To construct models for the identification of MCI and AD, we employed 8 machine learning algorithms, each selected based on specific features and suitability for MCI and AD prediction tasks. It's important to note that these machine learning algorithms were implemented using the Scikit-learn (sklearn)⁴³ framework for our research in this work.

Tree-Based Algorithms: Extra Trees (ET)³⁶: ET, by introducing more randomness, help improve model diversity to better capture differences between MCI and AD. Random Forest (RF)³³: RF has the potential advantage of capturing the relationships between MCI and AD, aiding in the discovery of critical diagnostic features.

Ensemble Learning Algorithms: Gradient Boosting (GB)³⁵ and Adaptive Boosting (ADA): These ensemble learning algorithms synthesize information from multiple weak classifiers, enabling the construction of more powerful models for accurate differentiation of MCI and AD.

Linear model

Logistic Regression (LR)³¹ is a linear classification model, which holds potential in distinguishing MCI patients and individuals at risk of progressing to AD, facilitating early intervention and treatment planning. Linear Discriminant Analysis (LDA)³⁷ maximizes the separation between classes by finding a linear combination of features while minimizing within-class variance. In the context of MCI and AD recognition, LDA enhances model performance and mitigates the risk of overfitting by reducing data dimensionality.

Support Vector Machine (SVM) ³⁸: SVM shows potential in delineating the boundary between MCI and AD by constructing an optimal hyperplane, helping clarify the differences between MCI and AD.

Gaussian Naive Bayes (GNB) ³²: GNB provides probability estimates for individuals with MCI and AD based on the normal distribution of features, aiding in identifying which features are particularly important for distinguishing MCI and AD. These algorithms were thoughtfully chosen for the distinct capabilities and suitability in addressing specific aspects of MCI and AD prediction tasks, enhancing the comprehensiveness and resilience of our models.

This work utilizes routine laboratory blood test data and machine-learning techniques to predict the risk of MCI converting to AD. Here, we first screened out 31 analytes shared by more than 90% of patients and excluded samples with any missing value of these variables. We next categorized the resultant samples into 4 groups (Female 65 ~ 74, Male 65 ~ 74, Female 75 ~ 89, and Male 75 ~ 89) to minimize the bias effect of gender and age. We identify a significant difference in 5 biomarkers (hemoglobin, HCT, neutrophils, RBC, and WBC) between MCI and AD patients among the 4 groups (Table 1). We subsequently used the 5 biomarkers combined with various machine-learning algorithms to train the models and found a relatively good accuracy in predicting AD conversion in the Female 65 ~ 74 group with the CV test set (AUC = 0.70) and independent test set (AUC = 0.76). We further used different feature selection methods along with different modeling methods to optimize the prediction accuracy in the other 3 groups and finally generated relatively accurate prediction models with an AUC of nearly 0.70. Our findings indicate the possibility of predicting AD risk in MCI subjects by constructing prediction models with routine laboratory blood test data and sophisticated machine-learning methods.

Although there were 210 original variables provided in the routine laboratory tests, most of them were excluded as they were not shared by more than 90% of patients. Ultimately, only 31 variables were included for modeling, which can be roughly divided into 4 groups: analytes related to liver function (ALT, ALP, Bilirubin), variables associated with renal function (Creatinine, Urea), minerals and proteins (Potassium, Sodium, Calcium, Phosphate, Albumin, Globulin, Protein), and blood cells (Eosinophil (absolute), Basophil (absolute), Neutrophil (absolute), Lymphocyte (absolute), Monocyte (absolute), Eosinophil (%), Basophil (%), Neutrophil (%), Lymphocyte (%), Monocyte (%), WBC, Platelet, RBC, MCV, MCH, MCHC, HCT, Hemoglobin, RDW). All these parameters have been reported to be associated with cognitive function or AD. For instance, plasma ALP was significantly higher in AD patients than in healthy controls and inversely correlated with cognitive functions⁴⁴. Urea levels were dramatically elevated in the AD brain, while potassium levels were significantly decreased^45,46. Compared with normal controls, AD patients had reduced lymphocytes and basophils, but elevated MCH and MCV¹⁴.

Consistent with previous studies, we found in this study that AD patients had lower levels of hemoglobin, HCT, and RBC, but higher levels of neutrophils and WBC than MCI subjects. Lower values of hemoglobin, HCT, and RBC suggest impaired oxygen supply or hypoxia in the brain, which could promote Aβ generation and neurodegeneration. It has been reported that decreased hemoglobin levels would facilitate neuroinflammation- and oxidative stress-mediated generation of Aβ to accelerate neurodegeneration^47,48. Meanwhile, Aβ impairs the morphology and function of RBC^49,50. Consequently, reduced hemoglobin, HCT, and RBC are closely associated with AD, and recovery of RBC number and quality might be beneficial to prevent AD. A significant increase in the population of neutrophils and WBC in AD patients suggests that neutrophil and leukocyte proliferation may be affected by AD-related pathological factors like neuroinflammation and oxidative stress reactions⁵¹. In fact, some inflammatory cytokines involved in AD pathogenesis like TNF-α and IL-9 can increase the neutrophils population^52,53. Elevated neutrophil counts, in turn, promote the activation of neutrophils and release of TNF-α by stimulating T cells⁵⁴. Moreover, activated neutrophils damage the blood-brain barrier, resulting in increased infiltration of inflammatory cells and cytokines in the brain, which exacerbates neurodegeneration^55–57.

Our study explores the potential of blood biomarkers in predicting AD risk among patients with MCI. Notably, our research diverges from prior studies⁵⁸ in several key aspects: Firstly, our study employs ensemble learning, which integrates results from multiple machine learning algorithms. In contrast, previous works solely utilized a single machine learning algorithm. This approach enhances the robustness of our model, providing a more comprehensive understanding of AD risk factors. Secondly, a pivotal aspect of our study lies in the choice of subjects. Unlike prior research that primarily concentrated on individuals with AD and normal cognition (NC), we specifically target individuals with AD and MCI. This shift is of paramount significance as MCI represents a transitional stage between NC decline and AD, making it a crucial area for early intervention and targeted therapeutic strategies. However, the study has limitations. Firstly, the age of the investigated subject is limited to 65 ~ 89 years old, so the prediction models cannot be generalized to the whole population. Secondly, useful information from participants is insufficient. For example, education level⁵⁹ has an important effect on AD, but it cannot be accessed because of the privacy-protection policy. The participants’ cognitive performance, such as the Mini-Mental State Examination (MMSE)⁶⁰ and the Montreal Cognitive Assessment (MoCA)⁶¹ scores, is not available either. Thirdly, the subjects were Chinese, and the models should be tested in other populations. In addition, the identification between MCI and AD is challenging, even using the specific plasma biomarkers (such as Aβ, tau, and phosphorylated tau) only achieves limited success²³, let alone the results of routine laboratory blood test are so general. Finally, even though screening is based on large sample size, the number of patients who can be trained is still limited, especially after a series of quality-controlling processes, which also reduces the model performance.

In the future, we will enlarge the sample size and collect more useful information to further optimize the model performances. For instance, we will supplement the results of plasma biomarkers like Aβ⁶², tau⁶³, and phosphorylated tau⁶⁴ when we enroll new participants, which are closely related to AD. We will also try to collect the cognitive evaluation of participants including MMSE and MoCA scores to improve the accuracy of models^60,65.

In this study, we identified 5 key routine blood biomarkers (Hemoglobin, HCT, Neutrophil (absolute), RBC, and WBC) which showed significant difference between MCI subjects and AD patients. We also established relatively accurate prediction models for MCI to AD conversion with routine blood test data and sophisticated machine-learning methods. Moreover, we developed an aid diagnostic platform (MAP, http://lab.malab.cn/~lijing/MAP.html) to predict the risk of AD progression in MCI subjects using routine blood test results. These findings provide a cost-effective and convenient method for the early screening of AD and help physicians focus on patients at high risk.

The study was approved by Institutional Review Board of The University of Hong Kong/Hospital Authority Hong Kong West Cluster (HKU/HA HKW IRB Ref No: UW 20-369).

ACKNOWLEDGEMENTS

The data is from the Hong Kong Hospital Authority Data Collaboration Laboratory (HADCL) (A000101). This study was supported in part by grants from the Innovation and Technology Commission – Hong Kong (MRP/056/21, PiH/234/22, InP/053/23 and PiH/086/22 received by Dr. You-qiang Song).

Tiwari S, Atluri V, Kaushik A, Yndart A, Nair M (2019) Alzheimer's disease: pathogenesis, diagnostics, and therapeutics. Int J Nanomed 14:5541–5554
Silva MVF, Loures CDG, Alves LCV, de Souza LC, Borges KBG, Carvalho MD (2019) Alzheimer's disease: risk factors and potentially protective measures. J Biomed Sci ;26
Barnes DE, Yaffe K (2011) The projected effect of risk factor reduction on Alzheimer's disease prevalence. Lancet Neurol 10(9):819–828
Jia JP, Wei CB, Chen SQ et al (2018) The cost of Alzheimer's disease in China and re-estimation of costs worldwide. Alzheimers Dement 14(4):483–491
Grassi M, Rouleaux N, Caldirola D et al (2019) A Novel Ensemble-Based Machine Learning Algorithm to Predict the Conversion From Mild Cognitive Impairment to Alzheimer's Disease Using Socio-Demographic Characteristics, Clinical Information, and Neuropsychological Measures. Front Neurol ;10
Petersen RC, Doody R, Kurz A et al (2001) Current concepts in mild cognitive impairment. Arch Neurol 58(12):1985–1992
Gomersall T, Smith SK, Blewett C, Astell A (2017) It's definitely Alzheimer's': Perceived benefits and drawbacks of a mild cognitive impairment diagnosis. Br J Health Psychol 22(4):786–804
Moradi E, Pepe A, Gaser C, Huttunen H, Tohka J, Initi ADN (2015) Machine learning framework for early MRI-based Alzheimer's conversion prediction in MCI subjects. NeuroImage 104:398–412
Zhang TT, Liao Q, Zhang DM et al (2021) Predicting MCI to AD Conversation Using Integrated sMRI and rs-fMRI Machine Learning and Graph Theory Approach. Front Aging Neurosci ;13
Syaifullah AH, Shiino A, Kitahara H, Ito R, Ishida M, Tanigaki K (2021) Machine Learning for Diagnosis of AD and Prediction of MCI Progression From Brain MRI Using Brain Anatomical Analysis Using Diffeomorphic Deformation. Front Neurol ;11
Minhas S, Khanum A, Alvi A et al (2021) Early MCI-to-AD Conversion Prediction Using Future Value Forecasting of Multimodal Features. Comput Intell Neurosci. ;2021
Chen YR, Qian XL, Zhang YY et al (2022) Prediction Models for Conversion From Mild Cognitive Impairment to Alzheimer's Disease: A Systematic Review and Meta-Analysis. Front Aging Neurosci ;14
Mapstone M, Cheema AK, Fiandaca MS et al (2014) Plasma phospholipids identify antecedent memory impairment in older adults. Nat Med 20(4):415–
Chen SH, Bu XL, Jin WS et al (2017) Altered peripheral profile of blood cells in Alzheimer disease A hospital-based case-control study. Medicine 96:21
Proitsi P, Kim M, Whiley L et al (2017) Association of blood lipids with Alzheimer's disease: A comprehensive lipidomics analysis. Alzheimers Dement 13(2):140–151
Boccardi V, Bubba V, Murasecco I et al (2021) Serum alkaline phosphatase is elevated and inversely correlated with cognitive functions in subjective cognitive decline: results from the ReGAl 2.0 project. Aging Clin Exp Res 33(3):603–609
Wu KY, Xu CY, Qiu GZ et al (2022) Association of lower liver function with cognitive impairment in the Shenzhen ageing-related disorder cohort in China. Front Aging Neurosci ;14
Baldini A, Greco A, Lomi M et al (2022) Blood Analytes as Biomarkers of Mechanisms Involved in Alzheimer's Disease Progression. Int J Mol Sci. ;23(21)
Gomes KB, Pereira RG, Braga AA et al (2023) Machine Learning-Based Routine Laboratory Tests Predict One-Year Cognitive and Functional Decline in a Population Aged 75 + Years. Brain Sci ;13(4)
Kononenko I (2001) Machine learning for medical diagnosis: history, state of the art and perspective. Artif Intell Med 23(1):89–109
Zhang F, Li Z, Zhang B, Du H, Wang B, Zhang X (2019) Multi-modal deep learning model for auxiliary diagnosis of Alzheimer’s disease. Neurocomputing 361:185–195
Yu K, Tan L, Lin L, Cheng X, Yi Z, Sato T (2021) Deep-learning-empowered breast cancer auxiliary diagnosis for 5GB remote E-health. IEEE Wirel Commun 28(3):54–61
Inoue M, Suzuki H, Meno K, Liu S, Korenaga T, Uchida K (2023) Identification of Plasma Proteins as Biomarkers for Mild Cognitive Impairment and Alzheimer’s Disease Using Liquid Chromatography–Tandem Mass Spectrometry. Int J Mol Sci 24(17):13064
Tanveer M, Richhariya B, Khan RU et al (2020) Machine learning techniques for the diagnosis of Alzheimer’s disease: A review. ACM Trans Multimedia Comput Commun Appl (TOMM) 16(1s):1–35
Falahati F, Westman E, Simmons A (2014) Multivariate data analysis and machine learning in Alzheimer's disease with a focus on structural magnetic resonance imaging. J Alzheimers Dis 41(3):685–708
Organization WH (2004) International Statistical Classification of Diseases and related health problems: Alphabetical index, vol 3. World Health Organization
Wan X, Wang W, Liu J, Tong T (2014) Estimating the sample mean and standard deviation from the sample size, median, range and/or interquartile range. BMC Med Res Methodol 14:1–13
Mishra P, Singh U, Pandey CM, Mishra P, Pandey G (2019) Application of student's t-test, analysis of variance, and covariance. Ann Card Anaesth 22(4):407
Branco P, Torgo L, Ribeiro RP (2016) A survey of predictive modeling on imbalanced domains. ACM Comput Surv (CSUR) 49(2):1–50
Haixiang G, Yijing L, Shang J, Mingyun G, Yuanyue H, Bing G (2017) Learning from class-imbalanced data: Review of methods and applications. Expert Syst Appl 73:220–239
Sagi O, Rokach L (2018) Ensemble learning: A survey. Wiley Interdisciplinary Reviews: Data Min Knowl Discovery 8(4):e1249
Browne MW (2000) Cross-validation methods. J Math Psychol 44(1):108–132
Wright RE (1995) Logistic regression
Ontivero-Ortega M, Lage-Castellanos A, Valente G, Goebel R, Valdes-Sosa M (2017) Fast Gaussian Naïve Bayes for searchlight classification analysis. NeuroImage 163:471–479
Rigatti SJ (2017) Random forest. J Insur Med 47(1):31–39
Meyer H, Reudenbach C, Hengl T, Katurji M, Nauss T (2018) Improving performance of spatio-temporal machine learning models using forward feature selection and target-oriented validation. Environ Model Softw 101:1–9
Friedman JH (2002) Stochastic gradient boosting. Comput Stat Data Anal 38(4):367–378
Sharaff A, Gupta H Extra-tree classifier with metaheuristics approach for email classification. Paper presented at: Advances in Computer Communication and Computational Sciences: Proceedings of IC4S 20182019
Balakrishnama S, Ganapathiraju A (1998) Linear discriminant analysis-a brief tutorial. Institute for Signal and information Processing. ;18(1998):1–8
Hearst MA, Dumais ST, Osuna E, Platt J, Scholkopf B (1998) Support vector machines. IEEE Intell Syst their Appl 13(4):18–28
He S, Guo F, Zou Q (2020) MRMD2. 0: a python tool for machine learning with feature ranking and reduction. Curr Bioinform 15(10):1213–1221
Shah C, Du Q, Xu Y (2022) Enhanced TabNet: Attentive Interpretable Tabular Learning for Hyperspectral Image Classification. Remote Sens 14(3):716
Kellett KAB, Williams J, Vardy ERLC, Smith AD, Hooper NM (2011) Plasma alkaline phosphatase is elevated in Alzheimer's disease and inversely correlates with cognitive function. Int J Mol Epidemiol Genet 2(2):114–121
Xu JS, Begley P, Church SJ et al (2016) Graded perturbations of metabolism in multiple regions of human brain in Alzheimer's disease: Snapshot of a pervasive metabolic disorder. Biochim Biophys Acta Mol Basis Dis 1862(6):1084–1092
Roberts BR, Doecke JD, Rembach A et al (2016) Rubidium and potassium levels are altered in Alzheimer's disease brain and blood but not in cerebrospinal fluid. Acta Neuropathol Com ;4
Peers C, Dallas ML, Boycott HE, Scragg JL, Pearson HA, Boyle JP (2009) Hypoxia and Neurodegeneration. Hypoxia Consequences Molecule Malady 1177:169–177
Salminen A, Kauppinen M, Kaamiranta K (2017) Hypoxia/ischemia activate processing of Amyloid Precursor Protein: impact of vascular dysfunction in the pathogenesis of Alzheimer's disease. J Neurochem 140(4):536–549
Jayakumar R, Kusiak JW, Chrest FJ et al (2003) Red cell perturbations by amyloid β-protein. Bba-Gen Subj 1622(1):20–28
Mohanty JG, Eckley DM, Williamson JD, Launer LJ, Rifkind JM (2008) Do red blood cell β-amyloid interactions alter oxygen delivery in Alzheimer's disease? Oxygen Transp Tissue Xxix 614:29–35
Pluta R, Ulamek-Koziol M, Januszewski S, Czuczwar SJ (2018) Platelets, lymphocytes and erythrocytes from Alzheimer's disease patients: the quest for blood cell-based biomarkers. Folia Neuropathol 56(1):14–20
Cowburn AS, Deighton J, Walmsey SR, Chilvers ER (2004) The survival effect of TNF-α in human neutrophils is mediated via NF-κB-dependent IL-8 release. Eur J Immunol 34(6):1733–1743
Holmes C, Cunningham C, Zotova E, Culliford D, Perry VH (2011) Proinflammatory cytokines, sickness behavior, and Alzheimer disease. Neurology 77(3):212–218
Dong XY, Nao JF, Shi JL, Zheng DM (2019) Predictive Value of Routine Peripheral Blood Biomarkers in Alzheimer's Disease. Front Aging Neurosci ;11
Baik SH, Cha MY, Hyun YM et al (2014) Migration of neutrophils targeting amyloid plaques in Alzheimer's disease mouse model. Neurobiol Aging 35(6):1286–1292
Pietronigro EC, Della Bianca V, Zenaro E, Constantin G (2017) NETosis in Alzheimer's Disease. Front Immunol ;8
Sayed A, Bahbah EI, Kamel S, Barreto GE, Ashraf GM, Elfil M (2020) The neutrophil-to-lymphocyte ratio in Alzheimer's disease: Current understanding and potential applications. J Neuroimmunol ;349
Stern Y, Gurland B, Tatemichi TK, Tang MX, Wilder D, Mayeux R (1994) Influence of education and occupation on the incidence of Alzheimer's disease. JAMA 271(13):1004–1010
Hoops S, Nazem S, Siderowf A et al (2009) Validity of the MoCA and MMSE in the detection of MCI and dementia in Parkinson disease. Neurology 73(21):1738–1745
Nasreddine ZS, Phillips NA, Bédirian V et al (2005) The Montreal Cognitive Assessment, MoCA: a brief screening tool for mild cognitive impairment. J Am Geriatr Soc 53(4):695–699
Zetterberg H, Wahlund L-O, Blennow K (2003) Cerebrospinal fluid markers for prediction of Alzheimer's disease. Neurosci Lett 352(1):67–69
Iqbal K, Liu F, Gong C-X, Grundke-Iqbal I (2010) Tau in Alzheimer disease and related tauopathies. Curr Alzheimer Res 7(8):656–664
Fossati S, Cejudo JR, Debure L et al (2019) Plasma tau complements CSF tau and P-tau in the diagnosis of Alzheimer's disease. Alzheimer's Dementia: Diagnosis Assess Disease Monit 11:483–492
Pinto TC, Machado L, Bulgacov TM et al (2019) Is the Montreal Cognitive Assessment (MoCA) screening superior to the Mini-Mental State Examination (MMSE) in the detection of mild cognitive impairment (MCI) and Alzheimer’s Disease (AD) in the elderly? Int Psychogeriatr 31(4):491–504

Table 1 is available in the Supplementary Files section.

There is NO Competing Interest.

Download PDF

Version 1

posted

You are reading this older preprint version

Read the latest preprint version →

An aid diagnostic platform to detect the transition of mild cognitive impairment (MCI) to Alzheimer's disease (AD) based on 48,116 AD and MCI patients

Status:

Version 1

Abstract

Figures

Introduction

Results

Individual biomarker analysis

Model development

Routine blood biomarker-driven models and effective predictive performance

Impact of age and gender on prediction performance

Comparative analysis of MAP and TabNet models for predicting the risk of MCI to AD

Methods

Participants

Data processing

Machine learning algorithm

Discussion

Conclusion

Declarations

ACKNOWLEDGEMENTS

References

Tables

Additional Declarations

Supplementary Files

Status:

Version 1