Alzheimer’s disease (AD) is the most common cause of dementia affecting the elderly (Jitsuishi and Yamaguchi, 2022; Sherimon et al., 2021; Zekri et al., 2015), and its incidence is expected to continue to increase as the population ages (Ivascu et al., 2015; Zhang et al., 2014). Mild Cognitive Impairment (MCI) has attracted a great deal of attention as a transitional stage between normal aging and AD (Jitsuishi and Yamaguchi, 2022; Panza et al., 2005; Zhang et al., 2014). Early detection of this stage is of vital importance for appropriate early intervention to help slow disease progression and improve patients’ quality of life (Ivascu et al., 2015; König et al., 2018). Therefore, significant efforts have been dedicated to identifying more efficiently early features and symptoms of MCI (Petersen et al., 2014) which has produced an exponential growth of biomedical data (Hoehndorf et al., 2015).

In recent years, ML techniques have been used to obtain an early diagnosis of MCI, either using MRI imaging (Jitsuishi and Yamaguchi, 2022) or neuropsychological tests (Clark et al., 2016; Linz et al., 2017; López-de-Ipiña et al., 2018) due to their capability of handling large amounts of information and obtaining clinically relevant knowledge (Sherimon et al., 2021; Weakley et al., 2015). But for this information to be useful, and the results obtained in studies with ML models to be generalizable, the data must be in a standardized format (Gomez-Valades et al., 2021; Sherimon et al., 2021; Zhang et al., 2014). This allows efficient retrieval of data (Patrick and Li, 2012; Sahoo et al., 2022), shareability between different centers (Gomez-Valadés et al., 2019), and univocal interpretation (Sherimon et al., 2021). Otherwise, the analyses could lead to different interpretations at centers other than where the data came from, or even in the same center because inexperienced staff may not be familiar with the original guidelines, or the population distribution changes over time. This is extremely critical in the healthcare field (Sherimon et al., 2021).

In this scenario, ontologies play a critical role in the management and interoperability of information, allowing the consistent representation of knowledge, standardizing data acquired and stored under different formats and protocols (Costa, 2014; Hoehndorf et al., 2015), providing a unique meaning to each element (Sherimon et al., 2021), avoiding interoperability problems (Gomez-Valadés et al., 2019; Kulmanov et al., 2020), easing the retrieval of information and records (Patrick and Li, 2012), and improving data analysis and efficiency of clinical diagnostic support systems (Shoaip et al., 2019).

Thus, on the one hand, ML models are used to obtain knowledge by searching for patterns of interest in large volumes of data (Tsymbal et al., 2007; Weakley et al., 2015), while, on the other hand, ontologies provide the basis for reusing and unambiguously integrating domain knowledge within applications (Jensen et al., 2013; Kang et al., 2019; Tsymbal et al., 2007). Our proposal seeks to leverage the benefits of both technologies, which separately have their inconveniences. In the case of ML models, it is usually difficult or even impossible to know the logical process behind a decision (Tsymbal et al., 2007; Weakley et al., 2015). Moreover, as they do not check data integrity, they can operate with conceptually but not technically incorrect data, leading to erroneous patterns when working with poorly curated databases (Sherimon et al., 2021). In the case of ontologies, a high-level representation for the formalization of knowledge (Tsymbal et al., 2007) can reach levels of abstraction and complexity that make their use impractical or not viable in real-world scenarios (Zekri et al., 2015). Although significant efforts have been made to combine both technologies (Kulmanov et al., 2021; Robinson and Haendel, 2020), methods that integrate them into decision support systems are still under development (Kulmanov et al., 2020).

In this paper, we propose to integrate a set of bootstrap aggregated (or bagged) decision trees for early diagnosis of MCI, which are represented as rules using the Semantic Web Rule Language (SWRL), with an already defined Ontology Web Language (OWL) ontology with specialized knowledge in neuropsychological tests, NIO (Gomez-Valades et al., 2021). In this way, the integrated system eases data standardization while providing a fast and interpretable first assessment of the cognitive status of subjects, saving physicians and neuropsychologists time and allowing them to reach a wider population during the screenings.

The rest of the paper continues as follows: Section 2 summarizes the state of the art of other approaches integrating ontologies and ML; Section 3 describes the methodology used to build the integrated system, detailing the ontology, the learning model, and the integration method, as well as the database; Section 4 details the performance results, compares them with other well-established ML models, and presents three use cases that show some advantages of this integration; Section 5 introduces the discussion of these results; and finally, Section 6 closes with the conclusions.

Ontologies and ML models conform the two main technologies for extracting, manipulating, and obtaining new knowledge within a domain (Kulmanov et al., 2020; Sahoo et al., 2022; Tsymbal et al., 2007). It seems logical that proper integration between them would result in an overall improvement in the performance of decision support systems (Kulmanov et al., 2020; Sahoo et al., 2022; Zhang et al., 2014). However, both technologies are usually employed separately (Tsymbal et al., 2007), although recently there has been an increased effort to combine them (Kang et al., 2019). This combination is performed following different objectives, such as the automatic completion of ontologies (Mežnar et al., 2022), the search for emerging knowledge in ontologies using ML techniques (Kulmanov et al., 2021; Robinson and Haendel, 2020), or the improvement of diagnosis in decision support systems.

Within this last group, some studies use ontologies and ML models sequentially: an ontology is first used to standardize and add semantic knowledge to a database, which is subsequently used to train the automatic system (Lakshmi et al., 2019; Sahoo et al., 2022; Tsymbal et al., 2007). Other studies focus on the integration of the predictive ML model in an ontology. Thus, a compact decision support system is generated. In this area, one of the first approaches appears in the work of Zhang et al. (2014) with Ontology-Driven Decision, which combines an ontology with a decision tree to create a decision support system for the early diagnosis of AD employing MRI images. In that system, the ontology is used to standardize the data and reduce subjectivity, while the decision tree generates the diagnosis. To integrate both parts, the decision tree rules were transformed into RDF rules, and the diagnosis was obtained using a reasoner. The work of Shoaip et al. (2021) proposes the integration of an existing ontology, ADDO (Shoaip et al., 2020), with a set of rules extracted from both a decision tree and a Repeated Incremental Pruning to Produce Error Reduction (RIPPER) method (Fürnkranz and Widmer, 1994) to differentiate between four categories (healthy, significant memory concern, early MCI, and late MCI). Unlike the other studies, which use their own database, they use a heterogeneous dataset obtained from the ADNI database (Petersen et al., 2010). This dataset includes neuropsychological tests, imaging tests, and chemical and genetic biomarkers which, together with the sociodemographic variables, are collectively called “biomarkers.” The rules obtained from ML were translated to SWRL rules to integrate them into the ontology. They also link different properties defined in the ontology using SWRL to give rules more expressiveness. Another approach also based on SWRL rules is the one proposed by Massari et al. (2022c) for diabetes detection, which combines a decision tree with ontologies. The diagnosis is obtained through inference using a reasoner. The system described was adapted afterward for the early detection of breast cancer (Massari et al., 2022a) and covid-19 (Massari et al., 2022b).

These papers showed that a proper integration of both, ontologies and ML algorithms, allows heterogeneous data to be accessed and put into a standardized framework, improving the performance of automatic models, and facilitating the exchange of data and results. However, there are still problems that need to be addressed. The first is the need to reuse or adapt previous ontologies in some studies that create their ontologies from scratch. Therefore, they may have redundancies and inconsistencies with prior ontologies. The second one is the selection of a decision tree or a RIPPER as the ML model, which was made to ease its translation and integration to rules (SWRL or RDF), something not possible with more complex ML models due to their black box structure. However, decision trees have strong training set dependency (Zhang et al., 2014), RIPPER methods have problems with noise and complex databases and need categorical data (Kotelnikov and Milov, 2018), and both methods are prone to overfitting if not pruned properly. Shoaip et al. (2021) mention using a decision tree and a RIPPER method to create SWRL rules but need to explain how they combine both sets of rules to function as one. Another problem is that of missing values. The most common ways to deal with missing values are either deleting the affected records or tests or imputing the missing values. However, there is an inherent risk of altering the database and hence the results.

Therefore, in this work we propose a decision support system that integrates a set of decision trees that work together as an ensemble to provide a diagnosis based solely on neuropsychological tests with an already established ontology, NIO. We used neuropsychological tests for being cheaper, faster, and less invasive than the alternatives while keeping a good diagnostic capability, making them the most suitable for population screenings. The ontological part will allow the data to be standardized and placed in a semantic context, and the tree ensemble will establish a diagnosis that combines the explainability of decision trees with the power and robustness of an ensemble method. The system is also designed to operate directly with databases with missing values without the need for prior preprocessing by deleting or imputing records.

The reliability of the integrated system in identifying cases with MCI was assessed by evaluating its performance in accurately classifying the cases. The versatility and ease of use of the system were also demonstrated through three additional practical use cases based on real-world scenarios: screening of new cases with the possibility of missing records, incorporation of a new database into the system, and the use of ontological capabilities to link different domains and generate new knowledge.

Owing to the operational mechanisms of the SWRL rules, the system meticulously records all intermediary decisions involved in deducing the final diagnosis from data. Its integration in an ontology allows the recovery of the inference process followed by the reasoner to reach a specific diagnosis. Using a decision tree ensemble instead of a single tree increases the system power, making it more accurate at identifying cases with MCI and less prone to overfitting the training set. Furthermore, the integrated system allows cases with missing data to be treated directly, without the need for prior data preprocessing. This avoids the possible inclusion of artifacts in the system.

Recall measures the model’s ability to retrieve positive samples (Gupta et al., 2021). Therefore, for our study, it is of great interest to increase recall to detect as many positive cases as possible. However, using recall alone without considering precision may result in a trivial model that would classify all subjects as MCI by default, while the goal of screenings is to reduce the number of subjects to focus on. F2 was selected as the most appropriate metric to evaluate the performance of a screening method as it gives more weight to recall while keeping a balance between FP and FN. Higher F-score was discarded as it could turn the screening system into a trivial one.

Two conclusions were drawn from the analysis of the tree ensemble thresholds. First, the most efficient threshold for a scenario in which both recall and precision would be optimized with 11 trees corresponds to th = 5. This means that a minimum of 5 decision trees would be necessary to issue a diagnosis of MCI and classify a case as such. Second, the tree ensemble with th = 2 is the most efficient detecting MCI cases, with F2_TE2 = 0.830 just after the BAG (F2_BAG = 0.832), as shown in Table 3. Therefore, the performance of the system is suitable to identify MCI subjects. It is worth noting that, in general, the thresholds in all machine learning models were low (from th = 0.2 to th = 0.4). This could indicate that when MCI symptoms are still mild, a low threshold would help detect a higher number of MCI cases that otherwise would have been classified as healthy.

The following can be inferred from the three use cases shown. In the first one, the integrated system can issue a diagnosis of new cases fast and without the need for preprocessing that could alter the data. The system also allows physicians and neuropsychologists to review each case and the inference followed if necessary. It is also possible to establish a diagnosis in those cases with missing data, even though the final diagnosis could not be obtained, using the “partial diagnosis” of the activated trees. The use of the original records, without the need for prior preprocessing to remove or impute missing values before analysis, allows experts to focus on the evaluation and diagnosis of the subjects. This shows the system’s ability to be used as a population screening tool, saving time in the diagnostic evaluation and, therefore, allowing more people to be reached.

In the second use case, the use of another database revealed the great heterogeneity present in the field of early detection of Alzheimer’s disease through neuropsychological tests. Because almost none of the tests presented in the Spanish and Norwegian databases matched, it was necessary to first check that all tests were modeled in the ontology. It was also necessary to generate a new tree ensemble model able to establish a classification from the new database. However, the process of creating and evaluating the new tree ensemble was faster as the entire system methodology was already defined. The performance analysis showed that the system with threshold th = 2 was among the best models when considering both F2 and recall for detecting MCI cases. Also, all methods present a clear improvement compared to the results with the Spanish database. Future analyses are needed to determine the reasons for the discrepancies in the performance of both databases.

The last use case showed how the ontology’s semantic relationship capabilities can be used to relate data from different domains. Specifically, to identify cognitive problems and potentially affected brain areas using their test performance. This complementary analysis can help to highlight the main cognitive problems of a subject regardless of his or her diagnosis. However, the information that links neuropsychological tests with both cognitive processes and brain areas is scarce, scattered, and often contradictory (Gomez-Valades et al., 2021), even in widely used and studied tests such as verbal fluency tests. Therefore, although an increase in this type of relationships may help to refine the MCI diagnosis, identifying the type of MCI and what kind of neurodegenerative disease is more likely to lead based on the test performance, in-depth modeling of this type of relationships is a project in its own, and it is left as future work.

This paper presents the design of a decision support system that integrates an ontology with a tree ensemble written under SWRL. The system allows the explainability of the generated diagnosis while maintaining performance on par with other well-established ML systems. Its ontological base allows the system to operate within the ontological framework: integrating the data in the ontology allows the standardization and univocal interpretation of the stored data, and defining value limits for each test minimizes the inclusion of tests with erroneous values. The use of a small tree ensemble to obtain the diagnosis allows us to combine the explainability and translation capacity of the decision tree with the power of a bagging method. Integrating it within the ontology allows a reasoner to explain the reasoning process. The use cases show its practical utility in three additional contexts: direct cognitive screening from a dataset without requiring previous preprocessing, such as the one obtained during real population screenings which can have missing values; integrating the necessary rules so that the system can generate diagnoses from information in a different database; and establishing relationships between different domains based on the performance of subjects in the tests.

It demonstrates the ability of the system to be used to perform a preliminary automatic diagnosis of subjects, using the available results obtained in the neuropsychological tests. The system is designed to filter out as many suspected cases of MCI as possible, allowing its use as an initial screening method in primary care units for older adults. We also showed its ability to be extended with new knowledge, and to employ semantic capabilities for inference of new knowledge.