Prediction and Monitoring of Progression of Alzheimer’s Disease: Multivariable approaches for decision support

Research output: ThesisDissertationCollection of Articles


Alzheimer’s disease (AD), the most common form of dementia, is a slowly progressing neurodegenerative disease, which cannot be cured yet. However, certain medications and lifestyle interventions can delay progression of the disease and its symptoms, thereby positively influencing both quality of life of patients as well as cost- effectiveness of healthcare. Early diagnosis of AD is important because such interventions should be started already at an early phase of the disease to have the best effect. However, early diagnosis is challenging because pathological changes in the brain occur years before the clinical symptoms become visible. In addition, the re- search during the past years has produced information from a large number of different tests and biomarkers that can potentially contribute to diagnosis and prognosis of AD. This excessive amount of data can cause information overload for clinicians, thus hampering the clinicians’ decision making. Data-driven analysis and visualization methods may help with interpretation and utilization of large amounts of heterogeneous patient data and support the clinicians’ decision-making process. Furthermore, the methods may aid in identifying suitable patients for clinical drug trials. The aim of the work described in this thesis was to develop and validate data- driven methods for predicting and monitoring progression of Alzheimer’s disease at the different phases of the disease spectrum, starting from normal cognition and ending to death, using data from neuropsychological and cognitive tests, magnetic resonance imaging (MRI), cerebrospinal fluid samples (CSF), comorbidities, and genetics (apolipoprotein E). The thesis consists of four original studies published as international journal articles. The first study focused on the early phase of AD. A supervised machine learning method called Disease State Index (DSI) was utilized to predict who of the individuals with subjective cognitive decline (SCD) will progress to a more severe condition, i.e., mild cognitive impairment (MCI) or dementia. The study population included 647 subjects from three different memory clinic-based cohorts in Europe. When all data modalities were combined, the area under the receiver operating characteristic curve (AUC) was 0.81 and balanced accuracy was 74%. Negative predictive value was high (93%), whereas positive predictive value was low (38%). Performance of the DSI method in terms of AUC decreased by 11% when validated with an in- dependent test set. Additional analyses suggested that several differences between the cohorts may explain the decrease in the performance. The second study focused on a more advanced disease stage. The DSI method was applied to longitudinal data collected from an MCI cohort of 273 subjects obtained from the Alzheimer’s Disease and Neuroimaging (ADNI 1) study. Longitudinal profiles of the DSI values differed between the subjects progressing to dementia due to AD and subjects remaining as MCI. In addition, two subgroups were found in the group remaining as MCI: one group with stable DSI values over time and another group with increasing DSI values, suggesting the latter group may progress to dementia due to AD in the future. This study also extended the Disease State Fingerprint (DSF) data visualization method for longitudinal data. The third study predicted hippocampal atrophy over 24 months using baseline data and penalized linear regression. The cohorts consisted of subjects with normal cognition, MCI, and dementia due to AD and were obtained from the ADNI 1 (n=530) and the Australian Imaging Biomarkers and Lifestyle Flagship Study of Ageing (AIBL, n=176) studies. The models including different data modalities per- formed better than the models including only MRI features. However, both models underestimated the real change at higher atrophy rate levels, the MRI-only models showing a greater underestimation. When predicting dichotomized outcome, i.e., fast vs. slow atrophy, the models obtained a prediction accuracy of 79-87%. The MRI-only models performed well when evaluated with an independent validation cohort (AIBL). The last study focused on the latest phase of AD by identifying which disease- related determinants are associated with mortality in patients with dementia due to AD. The cohort included 616 patients from the Amsterdam Dementia Cohort. Age- and sex-adjusted Cox proportional hazards models revealed that older age, male sex, and worse scores on cognitive functioning, as well as more severe medial temporal lobe and global cortical atrophy were associated with an increased risk of mortality. An optimal combination of variables comprised age, sex, performance on digit span backward test and Trail Making Test A, medial temporal lobe atrophy, and tau phosphorylated at threonine 181 in CSF. In conclusion, data-driven methods can be used for predicting and monitoring progression of AD from the mildest stages to the more advanced stages. Combining information from several data modalities provides better prediction performance than individual data modalities alone. The results also highlight the importance of the validation of the methods with independent validation cohorts. Introduction of these methods to different environments and countries may require harmonization of patient examination methods and diagnostic criteria.
Original languageEnglish
QualificationDoctor Degree
Awarding Institution
  • Tampere University
  • van Gils, Mark, Supervisor
  • Korhonen, Ilkka, Supervisor, External person
Award date17 Feb 2023
Print ISBNs978-952-03-2734-7
Electronic ISBNs978-952-03-2735-4
Publication statusPublished - 17 Feb 2023
MoE publication typeG5 Doctoral dissertation (article)


  • clinical decision support
  • data visualization
  • machine learning
  • multivariable modelling
  • prediction
  • supervised learning


Dive into the research topics of 'Prediction and Monitoring of Progression of Alzheimer’s Disease: Multivariable approaches for decision support'. Together they form a unique fingerprint.

Cite this