Data-Driven Differential Diagnosis of Dementia Using Multiclass Disease State Index Classifier

Antti Tolonen, Hanneke Rhodius‐Meester, Marie Bruun, Juha Koikkalainen, Frederik Barkhof, Afina W. Lemstra, Teddy Koene, Philip Scheltens, Charlotte E. Teunissen, Tong Tong, Ricardo Guerrero, Andreas Schuh, Christian Ledig, Marta Baroni, Daniel Rueckert, Hilkka Soininen, Anne M. Remes, Gunhild Waldemar, Steen Gregers Hasselbalch, Patrizia Mecocci, Wiesje M. van der Flier, Jyrki LötjönenFrontiers in Aging Neuroscience 2018 · DOI

Abstract

Clinical decision support systems (CDSSs) are increasingly vital for the differential diagnosis of neurodegenerative diseases. This study evaluates the PredictND tool, which utilizes a novel multiclass Disease State Index (DSI) classifier to differentiate between five cognitive states: subjective cognitive decline (SCD), Alzheimer’s disease (AD), vascular dementia (VaD), frontotemporal lobar degeneration (FTLD), and dementia with Lewy bodies (DLB). By integrating neuropsychological tests, MRI data, and cerebrospinal fluid (CSF) biomarkers, the classifier achieved a balanced accuracy of 82.3% in a cohort of 504 patients. Notably, accuracy rose to 93.6% for the half of patients where the system reported high confidence. The DSI classifier’s unique visualization of its decision-making process offers a transparent "fingerprint" for clinical use, bridging the gap between complex data and actionable medical insights.

Introduction: The Challenge of Differential Diagnosis

Differential diagnosis in dementia is a complex clinical task. While research often focuses on binary classifications (e.g., Alzheimer’s vs. Healthy Controls), real-world clinicians must distinguish between multiple overlapping etiologies. Patients frequently present with symptoms that could indicate AD, FTLD, DLB, or VaD simultaneously. Current diagnostic guidelines are often siloed by disease, and many existing machine learning models operate as "black boxes," providing predictions without interpretable reasoning—a significant barrier to clinical adoption.

Methods: The PredictND Tool and Multiclass DSI

The researchers developed a multiclass extension of the Disease State Index (DSI) classifier, integrated into the PredictND tool. Unlike standard algorithms, the DSI is built for interpretability. It functions through pairwise comparisons, computing a likelihood index (0 to 1) for every possible pair of diagnoses. These indices are then aggregated to determine the most likely class.

The PredictND tool interface showing the multiclass Disease State Index visualization and "fingerprint" analysis.

A key feature of the tool is the "Disease State Fingerprint," which visually represents how specific biomarkers—such as hippocampal atrophy or CSF tau levels—contribute to the final diagnosis. This allows clinicians to see exactly which data points support or conflict with a suggested disease state.

Dataset and Evaluation

The study utilized a subset of the Amsterdam Dementia Cohort, comprising 504 patients (average age 65 ± 8 years). The cohort included 118 controls with SCD, 223 patients with AD, 92 with FTLD, 47 with DLB, and 24 with VaD. This dataset represents a complex memory clinic population, providing a rigorous test for the classifier's ability to handle real-world diagnostic ambiguity.

Results: Performance and Data Synergy

The multiclass DSI achieved a balanced accuracy of 82.3% across all five classes, significantly outperforming standard ensemble learning algorithms like RUSBoost (75.5%). The system was most sensitive to Vascular Dementia (91.7%) and least sensitive to DLB (74.5%).

Classification performance metrics (Accuracy, Balanced Accuracy, and Sensitivity) for different combinations of neuropsychological (NP), cerebrospinal fluid (CSF), and MRI features.

Analysis of different feature sets revealed that automated MRI quantification (AMRI) was the strongest single data source (66.1% balanced accuracy). However, the highest performance was achieved through data synergy; combining neuropsychological tests with automated MRI increased accuracy to 77.1%, with the peak performance reached when all four data sources (NP, CSF, visual MRI, and automated MRI) were integrated.

Discussion and Clinical Implications

A critical finding of this study is the value of the classifier's confidence measure. In cases where the tool was most confident (the top 50% of the cohort), balanced accuracy soared to 93.6%. This suggests a tiered clinical workflow where the CDSS handles clear-cut cases, allowing specialists to focus their expertise on more ambiguous or "low-confidence" patients.

Despite its success, the study acknowledges the "ground truth" limitation: clinical diagnosis is not 100% accurate compared to post-mortem pathology. Furthermore, the current model assumes a single primary disease, whereas many elderly patients suffer from mixed dementia. Future work will focus on identifying co-morbidities and incorporating more specific biomarkers to improve the classification of challenging conditions like DLB.

Summary updated

Abstract

Clinical decision support systems (CDSSs) hold potential for the differential diagnosis of neurodegenerative diseases. We developed a novel CDSS, the PredictND tool, designed for differential diagnosis of different types of dementia. It combines information obtained from multiple diagnostic tests such as neuropsychological tests, MRI and cerebrospinal fluid samples. Here we evaluated how the classifier used in it performs in differentiating between controls with subjective cognitive decline, dementia due to Alzheimer's disease, vascular dementia, frontotemporal lobar degeneration and dementia with Lewy bodies. We used the multiclass Disease State Index classifier, which is the classifier used by the PredictND tool, to differentiate between controls and patients with the four different types of dementia. The multiclass Disease State Index classifier is an extension of a previously developed two-class Disease State Index classifier. As the two-class Disease State Index classifier, the multiclass Disease State Index classifier also offers a visualization of its decision making process, which makes it especially suitable for medical decision support where interpretability of the results is highly important. A subset of the Amsterdam Dementia cohort, consisting of 504 patients (age 65 ± 8 years, 44% females) with data from neuropsychological tests, cerebrospinal fluid samples and both automatic and visual MRI quantifications, was used for the evaluation. The Disease State Index classifier was highly accurate in separating the five classes from each other (balanced accuracy 82.3%). Accuracy was highest for vascular dementia and lowest for dementia with Lewy bodies. For the 50% of patients for which the classifier was most confident on the classification the balanced accuracy was 93.6%. Data-driven CDSSs can be of aid in differential diagnosis in clinical practice. The decision support system tested in this study was highly accurate in separating the different dementias and controls from each other. In addition to the predicted class, it also provides a confidence measure for the classification.
BibTeX
@inproceedings{tolonen2018data,
  author = {Antti Tolonen and Hanneke Rhodius‐Meester and Marie Bruun and Juha Koikkalainen and Frederik Barkhof and Afina W. Lemstra and Teddy Koene and Philip Scheltens and Charlotte E. Teunissen and Tong Tong and Ricardo Guerrero and Andreas Schuh and Christian Ledig and Marta Baroni and Daniel Rueckert and Hilkka Soininen and Anne M. Remes and Gunhild Waldemar and Steen Gregers Hasselbalch and Patrizia Mecocci and Wiesje M. van der Flier and Jyrki Lötjönen},
  title = {{Data-Driven Differential Diagnosis of Dementia Using Multiclass Disease State Index Classifier}},
  booktitle = {Frontiers in Aging Neuroscience},
  year = {2018},
  doi = {10.3389/fnagi.2018.00111},
}
Figure