IF-HUPM Algorithm Offers Interpretable Insights in Fuzzy Medical Databases

By Soham NandiReviewed by Susha Cheriyedath, M.Sc.Jan 16 2024

In an article published in the journal Nature, researchers introduced the improved fuzzy high-utility pattern mining (IF-HUPM) algorithm for computerized medical decision-making. Addressing the limitations of existing HUPM in medical databases, IF-HUPM enhanced data interpretability and explainability through fuzzy preprocessing.

Background

In the evolving landscape of artificial intelligence (AI) and medical diagnostic decision-making, leveraging computer-based methods has proven essential for accurate diagnoses. Current techniques, often rooted in statistics, machine learning (ML), and AI, mine vast datasets to extract patterns and associations for diverse applications, including recommendation systems. Traditional approaches like the Apriori algorithm and frequent pattern mining (FPM) focus on item set existence, lacking value consideration.

High-utility itemsets (HUI) mining addresses this by incorporating both frequency and item value. Despite its potential, HUI mining is primarily applied to profit-centric databases, neglecting interpretability. To bridge this gap, Fournier-Viger proposed an algorithm incorporating fuzzy set theory for F-HUPM. However, F-HUPM's interpretability in medical contexts remains a challenge.

This paper introduced the IF-HUPM algorithm, addressing the interpretability gap in existing efficient itemset mining models for medical data. IF-HUPM integrated fuzzy preprocessing to enhance interpretability in quantitative medical databases, utilizing fuzzy tree and list structures within an adaptive-phase framework.

By amalgamating one-phase and two-phase algorithmic properties, IF-HUPM ensured efficiency across diverse medical data properties. The proposed adaptive-phase fuzzy hybrid algorithm based on utility pattern growth and fuzzy hybrid mining (UFH) further refined F-HUPM based on multi-dimensional data fuzzification, presenting an efficient and interpretable decision-making framework for medical databases. Researchers contributed significantly to addressing the interpretability challenge in medical decision-making systems, promising advancements in intelligent decision-making and enhanced medical service quality.

F-HUPM algorithm and framework

The authors discussed the application of fuzzy set theory to the fuzzification process of medical databases, laying the groundwork for F-HUPM. Traditional HUPM algorithms employed either one-phase or two-phase approaches for discovering HUIs. The two-phase model utilized a tree-based structure for candidate generation and validation, incurring computational overhead. In contrast, one-phase models leveraged inverted lists, avoiding candidate set generation.

To integrate the advantages of both, a hybrid algorithm was proposed, which could dynamically select algorithm frameworks. However, defining criteria for transition remained a challenge. The authors outlined basic definitions crucial for understanding the F-HUPM process, including internal and external utility, transaction utility, support of item sets, and fuzzy set theory. They introduced the fuzzy HUI construction process, addressing challenges posed by floating-point medical data.

Fuzzification techniques, such as the fuzzy partition interval method, were discussed for transforming continuous data into boolean or categorical form. The proposed algorithm, IF-HUPM, focused on medical data, ensuring efficient and interpretable mining results, particularly in multidimensional medical scenarios. The algorithm considered medical reference values for human indicators, enhancing its applicability and interpretability in healthcare contexts.

The proposed algorithm

The IF-HUPM algorithm was introduced for extracting high-utility patterns (HUPs) from fuzzy medical databases. Addressing the challenge of choosing between one-stage and two-stage mining algorithms, the proposed adaptive-phase fuzzy UFH hybrid framework dynamically combined the advantages of fuzzy hybrid mining (FHM) and utility pattern growth. The algorithm scanned the dataset, constructed a fuzzy tree, and performed overestimated utility calculations, eliminating irrelevant itemsets. It then calculated actual utility values, identifying HUIs based on user-defined thresholds.

The algorithm integrated a switching module, enabling the transition from a tree-based to a list-based approach for improved efficiency. The adaptive-phase UFH hybrid framework accommodated both one-stage and two-stage mining algorithms, enhancing adaptability and efficiency.

The FHM algorithm within IF-HUPM utilized a vertical data mining approach, employing the utility-list data structure for extracting HUIs. FHM employed the utility list to identify HUIs, utilizing fuzzy sets for itemset evaluation. The algorithm organized the database based on itemset order, and a list of utilities was generated for each itemset. FHM did not generate candidate sets, reducing join operations using estimated utility co-occurrence pruning (EUCP). The proposed IF-HUPM algorithm combined the strengths of fuzzy set theory and HUPM, offering an efficient and interpretable solution for mining patterns in multidimensional medical data.

Experimental evaluation

The experimental evaluation compared the proposed IF-HUPM algorithm with the efficient HUI mining (EFIM) and UPHist algorithms, using diverse datasets including real and synthetic data. The datasets included the diabetes dataset, Pima dataset, foodmart dataset, and mushroom dataset, each with unique characteristics. The evaluation considered parameters such as runtime, memory usage, and the number of generated fuzzy HUIs across varying utility thresholds.

The results demonstrated the algorithm's performance on different datasets, highlighting variations in temporal and spatial complexities. The fuzzy UFH algorithm showed efficiency in pruning speed and memory scalability. The UPHist algorithm exhibited significant performance improvement, particularly at larger utility thresholds. The IF-HUPM algorithm struck a balance between execution time and memory consumption, showcasing its effectiveness.

Further evaluation based on the mushroom dataset confirmed the superior performance efficiency of the IF-HUPM algorithm, as it reported lower average temporal and spatial complexities. The final results presented fuzzy HUIs and their utility values, enabling the derivation of interpretable fuzzy rules for health indicators. The potential applications included deriving rules for health assessments based on HUPs. 

Conclusion

In conclusion, the proposed IF-HUPM algorithm combined one-phase and two-phase approaches, employing fuzzy processing for medical data. Demonstrating stability and efficiency, it outperformed EFIM and UPHist in time and space requirements. While improved comprehensibility was achieved through fuzzification, potential algorithmic complexity and data-type limitations existed.

Future research must focus on type-2 fuzzy set theory applications, enhancing F-HUPM interpretability for diverse scenarios, and exploring fuzzy association rules. The algorithm holds promise for multidimensional medical data challenges, with ongoing optimization and research anticipated to enhance its performance and applicability.