Comorbidity diagnosis using machine learning: Fuzzy decision-making approach

1 Introduction

Comorbidity, the simultaneous existence of multiple medical circumstances in a single patient, is a common healthcare challenge [1]. An accurate and timely diagnosis of comorbid diseases is essential for effective treatment and better patient outcomes. Several clinical tests are performed for early detection of comorbidities [2]. However, they are complicated diagnostic systems and rarely used, except for the significant risk of developing comorbidities. Early detection can provide faster recovery, thus reducing the consumption of resources, such as money and time [3].

With a proportion of 17% projected for 2035 compared to 8.9% in 2015, the frequency of people affected by more than four diseases is expected to nearly double. Furthermore, it has been observed that 2/3 of people experiencing this level of multimorbidity will grapple with mental health issues, such as depression, dementia, and cognitive impairment [4].

Artificial intelligence (AI) can provide a suitable solution in this field of study [5,6]. The future can be empowered by AI to treat patients with comorbidities and surpass the restrictions of siloed care. Through agile algorithms, personalized therapies can be weaved, presenting insights from the complicated tapestry of their interwoven circumstances. Healthcare systems will be revitalized by AI and their cost-saving potential through several benefits for the individual. It is challenging for clinicians to fully understand the demographic features related to comorbidity data and the complexity of detailed clinical information using traditional medical approaches. In addition, it is difficult to accurately predict and diagnose comorbidities [7]. Hence, AI is embraced to deliver optimal healthcare to patients with comorbidities. There are also some challenges for researchers in the field of AI in diagnosing comorbidities including availability of data (demographic and clinical), developing treatment plans, lack of comprehension of comorbidities, and monitoring patients remotely. Such challenges are likely to be overcome via the continued development of AI, which has evolved as a transformative solution to master the fast and precise analysis of complete datasets, such as clinical and demographic characteristics, and testing and diagnostic data. With this advanced diagnostic capability, AI has become a key player in improving the healthcare landscape of people with comorbidities. AI facilitates tailored care, which improves individualized medical treatment, especially for patients with comorbid diseases. Patients engage with online resources, support systems, and interactive platforms to receive customized recommendations and counsel, thereby improving patient care and quality of life. AI also helps reduce healthcare expenditures by creating an integrated health protection system. This paradigm reduces redundant confirmations and testing and streamlines healthcare operations to improve financial efficiency and accountability. Hence, AI applications are considered important in the field of health care [8].

Notable contributions have been made to machine learning (ML) as a subset of AI for the prediction and diagnosis of various diseases. In particular, supervised ML is a technique that includes training models on labeled data for classification or prediction [9].

Timely and accurate diagnosis is critical for effective treatment and improved patient outcomes. Traditional diagnostic methods often struggle with the complexity of comorbid conditions due to overlapping symptoms, leading to misdiagnosis, inappropriate treatment, increased healthcare costs, and higher mortality rates [10]. AI offers a promising solution by leveraging algorithms capable of analyzing complex patient data [11] to develop personalized treatment plans. Beyond diagnosis [12], AI can enhance patient quality of life through personalized care, online support, and improved medication management, empowering patients with informed decision-making capabilities and fostering collaborative healthcare practices [13,14].

For comorbidity disease diagnosis, supervised ML models have emerged as powerful devices, with the potential to improve efficiency and accuracy in healthcare decision-making. Within the context of comorbidity diagnosis, several ML models have been employed (Random Forest [RF], Naive Bayes [NB], K-Nearest Neighbor [KNN], and support vector machines [SVM]) [15,16]. However, the utilization of ML techniques to explore and predict comorbidity showed several challenges. These include data collection difficulties, the appropriate selection of algorithms, and model performance evaluations and interpretation.

In this study, standard ML techniques were used to model comorbidity disease diagnosis. Such models may encounter difficulties in the accurate detection of these relationships with limited data, such as the 271 cases available in this study [17]. Hence, owing to the potential limitations and inherent complexity of our dataset, it is more appropriate to utilize standard ML methods. Ensuring that the model is efficiently developed and deployed, this choice prioritizes implementation speed. Moreover, satisfactory performance is often achieved using ML models at lower computational costs than their deep learning counterparts [18].

Regardless of such attempts, challenges still exist, as accuracy remains a concern in diagnosing comorbidity and suboptimal data selection is based on the affected properties. Therefore, ML methods can assist in the diagnosis of comorbidities, mainly by using datasets tailored specifically to such domains [19].

Fuzzy decision by opinion score method (FDOSM) and AI technologies have emerged in the clinical domain, as pivotal contributors to intelligent decision-making, particularly in disease detection and diagnosis. The measurement and evaluation processes are complex because of multi-criteria decision-making (MCDM). This can be solved by the present contribution proposing a novel dynamic framework designed to assess and benchmark crossover or hybrid diagnostic models for multiple diseases. In such innovation, it is essential to develop a diagnostic model at the intersection of ML algorithms and feature selection (FS) approaches, presenting a multitude of ML models for comprehensive evaluation. The main contributions of this article are as follows:

• Hybrid diagnosis model is proposed based on the intersection between ML and FS techniques for the detection of comorbidities.

• A dynamic decision-making system is developed to evaluate and benchmark the proposed hybrid mode based on various performance matrix criteria such as precision, recall, F1-score, Kappa, and Jaccard.

• Weighting the effective criteria of the proposed hybrid model for disease diagnosis and selecting the best available hybrid model using FDOSM.

The remainder of this article is organized as follows. Section 2 presents a literature review, while Section 3 shows the proposed model with concept of the FDOSM, and Section 4 presents the results and discussion. Finally, Section 5 presents conclusions and future work.

2 Literature review

In this literature review, key advancements and studies are considered. Exemplary studies pertinent to this domain include that introduce an FS algorithm for predicting comorbidity risk based on diagnosis codes in patients through supervised ML techniques. A systematic review was presented in the study of Jovel and Greiner [20], revealing 33 ML models with 80% accuracy in comorbidity prediction.

AI demonstrates outstanding effectiveness in the prediction of latent risks related to depression and the environment of comorbidities. Sánchez-Rico and Alvarado [21] suggested a model trained on data sourced from over 13,270 patients with a commendable accuracy rate of 95% in the prediction of comorbidities. An AI-based model tailored to assess comorbidity risk was developed by a team of university researchers. Training was performed on data encompassing more than 50,000 patients, achieving a considerable accuracy rate of 85% in predicting comorbidities (in relation to COVID-19) [22]. In recent research, the vast potential of AI has been underscored in revolutionization of comorbidity diagnosis. It was demonstrated that AI is effective in predicting comorbidities across diverse diagnoses, thus revealing its extensive applicability. Such an exciting trend is extended to the realm of HIV care leveraging ML techniques to investigate electronic health record (EHR) data and evaluate comorbidity patterns comprehensively [23]. AI can accurately analyze extensive datasets such as medical imaging, genetic data, and healthcare records as an appreciated tool for healthcare professionals to identify patients at risk of comorbidities [24], XGBoost was used [25] along with the explainability method for the prediction of comorbid conditions for chronic diseases such as heart disease and chronic obstructive pulmonary disease (COPD). Patients with higher risk or urgent-care-needing congenital heart disease (CHD) are more likely to experience complications or die. Care for such patients can be improved through classification and prioritization processes [26].

In the study of Shrot et al. [27], the comorbidities of a genetically rare condition were addressed. People diagnosed with psychogenic non-epileptic seizure (PNES) within the EMU were concentrated through previously examined data. Massot-Tarrús et al. [17] tried to recognize baseline characteristics and distinguish PNES patients only from those with comorbid epilepsy. Hence, it is important to understand the models rather than the outcomes of selecting an ML model. In conclusion, diagnostic considerations toward the significance of overlapping features and symptoms were neglected along with their effects and priorities in the lives of patients with comorbid diseases. The imperative for continuous development efforts and research is to focus on a diverse array of methodologies to standardize and improve these models for therapeutic applicability. Ensuring the trustworthiness of AI in comorbidity diagnosis is paramount, as medical decisions in terms of flawed AI outcomes could detrimentally affect patient well-being. Thus, standardized metrics and methodologies are required to evaluate the reliability and accuracy of AI-based healthcare systems and make well-informed medical judgments and, protection to patient health. It should be noted that, as far as the authors are concerned, no evaluation framework has been presented in the literature to benchmark the best hybrid diagnosis models and classify multiple diseases, particularly comorbidities. Therefore, multi-criteria evaluation solutions should be explored in the clinical domain. Therefore, researchers should work on reliable datasets.

In this study, we used a dataset published by the London Health Sciences Centre, which shows the raw dataset of patients diagnosed with PNES in the epilepsy monitoring unit (EMU). This dataset provides a demographic and clinical characterization of patients with PNES and a good model of comorbidity [17]. Furthermore, the literature highlights the need for an evaluation framework to benchmark hybrid diagnosis models used for multiple diseases, especially comorbidity, and multi-criteria evaluation and fuzzy MCDM techniques are proposed to intervene in clinical fields to provide intelligent decision-making, particularly for disease diagnosis and detection. The process of evaluation and benchmarking falls under the MCDM problem. Therefore, this study aims to develop a new dynamic framework for evaluating and benchmarking hybrid diagnosis models for multiple diseases.

3 Proposed model

The methodology used in this study, included six successive phases. In Phases 1 and 2, the data were identified and pre-processed. Then, diagnostic model is developed, which entails the intersection of the two prominent algorithms of FS (Chi 2, KBest) with 11 distinct ML algorithms (these processes are represented in Phases 3 and 4). The evaluation and measurement framework were included in Phases 5 and 6 using the FDSM technique. The methodological approach used in this study is illustrated in Figure 1.

Figure 1

The methodology of evaluation and benchmarking comorbidity diagnosis models. Source: Created by the authors.

3.2 Phase 2: Pre-processing

The comorbidities dataset included several sources (public and private). The data were not limited to comorbidity and trauma with chronic diseases. Data pre-processing involves the removal of missing values, data coding and cleaning, dataset normalization, and data imbalance.

3.2.1 Data coding and cleaning

A fundamental pre-processing procedure is represented by data cleansing, which can be used for any dataset to improve data quality, reliability, and accuracy in subsequent analyses. The meticulous handling of unidentified symbols and the identification of outliers have emerged as imperative tasks within the context of the comorbidity dataset, which is pivotal for reinforcing data integrity and cultivating insightful interpretations. The removal of extraneous symbols is undertaken rigorously in the process of data cleansing, which is not limited to characters such as “&,” “?,” “/,” and “-.” In this critical step, the dataset is streamlined, making it more conducive to comprehensive modeling and analysis [28]. Finally, owing to the pivotal role of a strong data-cleansing procedure in refining the comorbidity dataset, its suitability is guaranteed for subsequent ML applications. The precision and efficiency of the analytical process are improved by the removal of superfluous symbols and transformation of textual data into numeric formats.

3.2.2 Handling missing values

Missing values were included in the comorbidity dataset used in this study (Figure 2). Various methods can be used to address this issue by imputing and manipulating missing values. These approaches include model-based imputation (simple tree), random value imputation, distinct value imputation, and mean imputation, which are the most commonly used approaches in the following equation:

(1) Mean = Sum ( z i ) Count ( z i ) ,

where Z i represents the number of observed values for feature i .

Figure 2

The percentage of missing values in the comorbidity dataset. Source: Created by the authors.

3.2.3 Dataset normalization

For normalization, the raw data were either modified or re-scaled. The Min-Max normalization approach was used in this study for the comorbidity dataset in the following equation:

(2) x n = x − min ( x ) max ( x ) − min ( x ) ,

where x n represents the normalized feature, while x is the original feature.

3.2.4 Data imbalance

It is worth noting that data asymmetry is overlooked by some researchers making the models biased; as such, accuracy cannot represent the model integrity. An imbalance was exhibited by the dataset used in this study, with adverse effects on the accuracy of the diagnosis process. The synthetic minority oversampling technique (SMOTE) method is used to mitigate this issue because a resampling method is frequently utilized in ML to balance data in terms of the target class. By utilizing SMOTE, a higher accuracy can be obtained using the developed ML models in classification, thus providing a more realistic perception of the data.

3.6 Phase 6: FDOSM

The FDOSM is a fuzzy technique presented in this study for MCDM [31] and serves as a benchmark for the dimension reduction model. Three principal stages are included in the FDOSM. First, the input unit is intricately linked to the decision matrix (DM), as detailed in Table 3. Second, the data transformation unit as the center stage, facilitates the transfer of the DM into an opinion DM. Finally, within this unit, the conversion of the opinion DM is executed into a fuzzy opinion matrix by the data processing unit using a Likert scale (i.e., the value that would be provided to the model). Subsequently, using direct aggregation, the final rank is established for accessible alternatives (i.e., the model). Both units of the FDOSM are shown in Figure 3. It is essential to benchmark the developed models to choose the best model in terms of the performance evaluation metrics (criteria).

Table 3

Decision matrix

Alternatives			Performance evaluation metrics criteria
Diagnosis models			C1	C2	C3	C4	C5	C6	C7	C8
A1	KBest-DT	A12	Chi 2-DT	C1–Ai	C2–Ai	C3–Ai	C4–Ai	C5–Ai	C6–Ai	C7–Ai	C8–Ai
A2	KBest-RF	A13	Chi 2-RF	C1–Ai	C2–Ai	C3–Ai	C4–Ai	C5–Ai	C6–Ai	C7–Ai	C8–Ai
A3	KBest-GB	A14	Chi 2-GB	C1–Ai	C2–Ai	C3–Ai	C4–Ai	C5–Ai	C6–Ai	C7–Ai	C8–Ai
A4	KBest-AdaBoost	A15	Chi 2-AdaBoost	C1–Ai	C2–Ai	C3–Ai	C4–Ai	C5–Ai	C6–Ai	C7–Ai	C8–Ai
A5	KBest-Bagging	A16	Chi 2-Bagging	C1–Ai	C2–Ai	C3–Ai	C4–Ai	C5–Ai	C6–Ai	C7–Ai	C8–Ai
A6	KBest-ET	A17	Chi 2-ET	C1–Ai	C2–Ai	C3–Ai	C4–Ai	C5–Ai	C6–Ai	C7–Ai	C8–Ai
A7	KBest-KNN	A18	Chi 2-KNN	C1–Ai	C2–Ai	C3–Ai	C4–Ai	C5–Ai	C6–Ai	C7–Ai	C8–Ai
A8	KBest-SVM	A19	Chi 2-SVM	C1–Ai	C2–Ai	C3–Ai	C4–Ai	C5–Ai	C6–Ai	C7–Ai	C8–Ai
A9	KBest-SGD	A20	Chi 2-SGD	C1–Ai	C2–Ai	C3–Ai	C4–Ai	C5–Ai	C6–Ai	C7–Ai	C8–Ai
A10	KBest-Ridge classifier	A21	Chi 2-Ridge classifier	C1–Ai	C2–Ai	C3–Ai	C4–Ai	C5–Ai	C6–Ai	C7–Ai	C8–Ai
A11	KBest-ANN	A22	Chi 2-ANN	C1–Ai	C2–Ai	C3–Ai	C4–Ai	C5–Ai	C6–Ai	C7–Ai	C8–Ai

C = Criteria, A = Alternative, C1 = Precision, C2 = Recall, C3 = F1-score, C4 = Kappa, C5 = Jaccard, C6 = Hamming loss, C7 = MCC, C8 = Accuracy, i = 1–22.

Figure 3

FDOSM stages. Source: Created by the authors.

3.6.1 Input unit

The Data Entry Module is included in the first stage and is closely linked to the DM. This module creates potential solutions (hybrid models A 1 , … , A m and decision criteria sets C 1 , … , C n ) to select the best ML algorithm. The output (DM) is a matrix of possible options that is further transformed into a reasonable opinion matrix.

Dynamic DM was used to assess and benchmark the diagnostic models. Recognized as the paramount element in benchmarking and evaluation methodology [26,32], decision-making is based on the interplay between alternatives and choice criteria. The criteria for assessment serve as metrics to benchmark the 11 diagnostic models, symbolizing the alternatives array. Table 3 lists the intricate steps undertaken to create the DM.

Furthermore, the issues are conflicted and traded-off via the evaluation results based on the criteria to determine the best model. Hence, these issues are solved using the benchmarking and evaluation through FDOSM.

3.6.2 Data transformation module

The FDOSM data transformation module is the second stage, with a pivotal role during DM generation in selecting the optimal three-parameter solution (maximum, minimum, and critical values). For the cost criteria, minimum values are considered, and the most advantageous option is signified by the lowest value. Conversely, for benefit criteria, the most beneficial solution had the highest value. The critical value is applied when the optimal response is neither the maximum nor minimum. The message model is then transformed into an opinion matrix, as follows:

Step 1: For each criterion, to determine the optimum solution in the decision matrix, equation (13) is used.

(13) A = { [ ( max v i j J ε j ) . ( min v i j ∣ j ε J ) . ( O p i j ε 1 . J ) i = 1 , 2 , 3 , … , m ] } .

Step 2: Ideal solutions and other values are compared for each criterion by using linguistic terms in the benchmarking process. Five scales were used: “no difference,” “a slight difference,” “a difference,” “a large difference,” and “a very large difference.” The value is compared with alternative subsequent steps after selecting the optimal solution based on the same criterion equation:

(14) OP Long = ( ( U ˆ i j U i j ∈ j ) , i = 1 , 2 , 3 , m ) .

3.6.3 Data-processing unit

The opinion matrix represents the output of the transformation unit. The opinion matrix is subjected to the fuzzy value (FV) method to transform it into a fuzzy decision matrix, which initiates the final block. The next step of the procedure uses a direct aggregation operator such as the arithmetic mean [33]. Table 4 shows how the linguistic phrase was transformed into FV. This process was performed after comparing the optimal solution with various DM levels. The order with the highest connection to being the optimal ranking order is probably the one with the lowest mean score value.

Table 4

Priority and FV

Priority with acronyms	Fuzzy value
Very high (VH)	0.00–0.10–0.30
High (H)	0.10–0.30–0.50
Low (L)	0.30–0.50–0.75
No (N)	0.50–0.75–0.90

4 Results and discussion

In this section, the results of each phase of the study are presented. The results of the proposed models incorporating ML algorithms and FS to identify comorbid disorders are presented. We gathered and pre-processed the PNES dataset to accurately diagnose and predict the presence of comorbid disorders and to create a decision matrix. The weighted output is inserted into the FDOSM for benchmarking.

The outcomes extracted from the suggested models comprise solely ML algorithms to discern the pathological disorders associated with comorbidities. Tables 5, 6, 7 display the performance metrics outcomes for the 11 diagnosis models evaluated using the developed DM. Seventy percentage of the dataset was used in the evaluation of these models for training and the remaining 30% was used for testing.

After extracting the criteria for all features using various evaluation metrics, we utilized the Chi 2 algorithm with different FSs. Hence, an accuracy rate of 91.463% was achieved (using the Gradient Boost and AdaBoost). For both models, this surpasses the accuracy obtained by competing models for the diagnosis of a disease. We utilized (KBest) with various features, and a higher accuracy was obtained than the Gradient Boost algorithm, with an accuracy rate of 92.683%.

Tables 8, 9, 10 provide a comprehensive overview of the evaluation of the 11 models, considered as alternatives, across 8 definite performance assessment metrics, each functioning as a criterion. Weighted datasets were incorporated into the evaluation procedure. It should be noted that the FDOSM technique is effective as a guiding mechanism in using algorithms for the comorbidity dataset, leading to considerable enhancements in the performance metric values.

Table 11 presents the ranking outcomes for the 11 models, such as score values and orders to present a clearer perspective of their relative performance. Based on the FDOSM concept, the most preferable alternative is signified by the option with the lowest score, whereas the least desirable choice is represented by the option with the maximum score. Remarkably, the ultimate rankings and scores were reported to contextualize the perspective of DM. According to the tabulated data, the ET algorithm emerges as the optimal choice based on the preferences, when all features are used. In contrast, the AdaBoost algorithm was designated using the Chi 2 algorithm as the most effective ML technique to obtain a top score of 1. In this scenario, KNN ranked the ML method with the least effectiveness with a score of 10. Using the KBest algorithm, Gradient Boost outperforms with a score of 1, whereas the least favorable performance is recorded by SVM, securing a score of 11. As illustrated in Table 11, which shows the final decision (FD), AdaBoost appears as the top-performing algorithm. The ET algorithm is closely based on the second position as shown in Figure 4.

Figure 4

Plot shows rank and score. Source: Created by the authors.

However, this study has some limitations. First, only data from the London Health Sciences Center (LHSC, London, Canada) were used. This study must be replicated using data from other healthcare systems. The main limitation of this study is its small sample size. However, other features (160) were used to characterize it. The accuracy of the results can be improved using a larger dataset. Further studies are required to enhance the efficiency and accuracy of the AI methods utilized in comorbidity diagnosis. Furthermore, the protection of ML models is essential for vulnerabilities and threats, particularly adversarial attacks. Manipulation of the input data is included in adversarial attacks to fool the model, resulting in imprecise predictions in comorbidity diagnosis. ML models of comorbidities are susceptible to manipulation owing to their reliance on different features, such as demographics, imaging, and laboratory results. Thus, robust models must be developed by our research endeavors with resistance to such attacks using methods such as adversarial training. In addition, we should design a trustworthy framework to ensure reliable measurement and evaluation of these models.

AdaBoost has the highest effectiveness, scoring (0.154166667) and ranking first by leveraging weak classifiers to enhance predictive accuracy. The extra tree closely followed a score of (0.258333333), demonstrating robust performance in ensemble learning. RF and Bagging share the third rank with scores of (0.279166667), showing their ability to improve accuracy through ensemble methods. Gradient Boost ranked fifth with a score of (0.291666667), effectively refining predictions through gradient descent. DT followed with a score of (0.3), while SGD and ANN scored (0.408333333) and (0.435416667), respectively, showing moderate effectiveness. The SVM, KNN, and Ridge Classifier face challenges with scores of (0.4625), (0.514583333), and (0.616666667), respectively, indicating difficulties in leveraging the selected features effectively. Ensemble methods such as AdaBoost and ET prove advantageous for accurate modeling tasks with Chi 2 selected features, whereas more complex models may require additional tuning or alternative FS approaches to improve performance.

Gradient Boost leads with the highest effectiveness, scoring (0.25) and securing the top rank by iteratively improving model weaknesses through gradient descent. DT, RF, Bagging, and ET share the second rank with scores of (0.3), demonstrating robust performance in utilizing the selected features, which was achieved because of the results illustrated in Table 6, in which only two features are selected based on KBest. It should be noted that because these algorithms belong to the same DT family, the results are similar. AdaBoost follows with a score of (0.30625), ranking sixth, leveraging weak classifiers to achieve strong overall accuracy. KNN and SGD share the eighth rank with scores of (0.433333333), indicating moderate effectiveness. ANN ranks seventh with a score of (0.38125), which is effective in capturing complex patterns but require careful tuning. the ridge Classifier and SVM ranked 10th and 11th, respectively, with scores indicating challenges in leveraging the selected features effectively. Ensemble methods such as Gradient Boost and DT are advantageous for accurate modeling tasks with KBest features, whereas more complex models may require additional optimization to enhance performance.

This is calculated by averaging the various quality measures, including all features, chi-square, and KBest. AdaBoost emerged as the top performer with a score of (0.204861111), showcasing its effective combination of weak learners to enhance predictive capability. ET achieves a score of (0.230555555), highlighting its robust performance in ensemble learning with randomized splits. Bagging and RF secure 3rd and 4th ranks, respectively, with scores of (0.274305556) and (0.30625), leveraging bootstrap aggregation and ensemble methods to improve accuracy. Gradient Boost ranked fifth with a score of (0.326388889), refining predictions through gradient descent. DT, although interpretable, ranks sixth with a score of (0.345138889), showing slightly lower accuracy compared to ensemble methods. Furthermore, the ranks, ANN, SGD, Ridge Classifier, KNN, and SVM demonstrated varied performance, with SVM scoring the highest at (0.593055556), indicating challenges in handling complex data. As indicated previously, ensemble methods such as AdaBoost and Extra Tree stand out as optimal choices for tasks requiring robust and accurate ML models.

The performance of the best ML algorithms across various FS methods including all features, chi-square, and KBest ultimately determines their effectiveness through an FD score and rank. AdaBoost consistently emerged as the top performer, achieving the lowest scores across all FS methods with scores of (0.154166667) to (0.30625), resulting in a final score of (0.204861111) and ranking first overall. The extra Tree also demonstrates strong performance initially with a score of (0.133333333), maintaining a second rank consistently across FS. Gradient Boost shows robustness with scores ranging from (0.25) to (0.4375), securing a fifth rank in the FD. The results underscore AdaBoost and Extra Tree as reliable choices for accurate modeling across various feature subsets, while emphasizing the impact of FS methods on algorithm performance and ranking.

The following are the recommendations of this study when using a small dataset with ML algorithms:

• Ensemble methods such as ET, AdaBoost, and Bagging dominate the top ranks, showcasing their ability to improve predictive performance by combining multiple models.

• AdaBoost is the top-performing algorithm across all FS methods, showing its robustness and consistency in enhancing the prediction accuracy.

• More complex models such as ANN and Gradient Boost, while powerful, do not necessarily outperform simpler ensemble methods in this evaluation. This highlights the importance of model selection and tuning based on specific datasets and problems.

• Simpler models, such as DT and ridge classifiers perform moderately well, suggesting that while complexity can help, simplicity with the right approach (e.g., ensemble techniques) can often yield competitive results.

• More complex models such as SVM, ANN, and ridge classifiers show higher scores and indicate difficulties in effectively utilizing the selected features compared to simpler algorithms. This could be the result of utilizing a small dataset.

The integration of AI and fuzzy decision-making into real-time comorbidity diagnosis presents a transformative potential for healthcare. Here are some key implications and future applications: (i) handling uncertainty, (ii) personalized medicine, (iii) enhanced efficiency, and (iv) equitable care delivery. Despite this, various limitations, including the performance of the algorithms may vary with different datasets or additional FS techniques. Moreover, the FD score is an aggregated measure, which might obscure nuanced performance details. Furthermore, the proposed model was not tested on adversarial attacks to verify the generalization ability of the ML algorithms.

5 Conclusion and future work

In this article, we have used 11 distinct AI models with the advanced FDOSM tool for model selection to tackle the challenging problem of detecting comorbidities. Extensive demographic and clinical data are assimilated by AI, leveraging this knowledge to evaluate an individual’s risk of multiple diseases, even those not associated conventionally. Through such early detection capability, physicians are provided with valuable insights to prompt treatment initiation and intervention, thus presenting considerable implications for patient health. Our proposed technique illustrated the power of FDOSM as a model selection tool for the detection of comorbidities. This could allow the AI models to improve diagnosis and refinement of the predictions to tailor treatments with greater precision in which the simulation results showed 91.463% accuracy using AdaBoost technique. However, some of the limitations involving performance variance with different datasets or additional FS techniques. Furthermore, the FD score is an aggregated measure which might obscure nuanced performance details. Future work could involve the use of other FS techniques such as genetic algorithms and gray wolf techniques. Another direction of research is the use of federated learning for training the system to provide security and data integrity.