Applicability of Artificial Intelligence in the Field of Clinical Lipidology: A Narrative Review

Since epidemiological studies first established total cholesterol and low-density lipoprotein cholesterol (LDL-C) as causative factors in the development of atherosclerotic vascular disease, there has been an exponential increase in research in the field of lipidology.1

In recent years, artificial intelligence (AI) technologies have been transforming medicine and healthcare.2 AI refers to the simulation of human intelligence processes by machines, particularly computer systems. Machine learning (ML), a subcategory of AI, enables computers to identify patterns in vast datasets and formulate responses to various questions across multiple disciplines through the use of algorithms. Furthermore, deep learning (DL) is a subset of ML distinguished by its use of biologically inspired neural network architectures to process inputs.

The advanced technologies developed for AI have fueled its applications in many areas, such as security, business administration, finance, commerce, tourism, meteorology, and biology, among others.3 Similarly, AI's applications in the health sciences, especially in medicine, are extensive and span nearly all domains, such as genomics, epidemiology, diagnostics, prognosis, clinician workflow, telemedicine, and administrative management.4 Furthermore, the remarkable expansion of clinical lipidology in recent years has seen numerous intersections with AI. The pursuit of precision medicine within certain contexts of contemporary lipidology may benefit from these innovative techniques.

Therefore, the aim of this narrative review was to explore some of the applications of AI in the field of clinical lipidology.

A literature search was conducted to identify studies that established connections between various aspects of lipidology and AI. Two independent reviewers systematically searched the electronic databases of PubMed/MEDLINE, Embase, Science Direct, Scopus, and SciELO. The search terms used included “artificial intelligence,” “deep learning,” and “machine learning,” which were combined with the following lipid-related terms: “cholesterol,” “dyslipidemia,” “low-density lipoprotein cholesterol,” “high-density lipoprotein cholesterol,” “triglycerides,” “hypercholesterolemia,” “lipid-lowering treatment,” and “statins.” The search for relevant articles concluded on July 31, 2023.

Studies that analyzed the applicability of AI techniques for different diagnostic, prognostic, or therapeutic aspects of lipidology were included. There were no restrictions based on language, geography, or publication status. Studies that were excluded consisted of expert opinions, reviews, and those with a patient population of fewer than 100.

FH is the most common genetic disorder of lipid metabolism. FH results in life-long exposure to high LDL-C levels, which, if left untreated, significantly increase the risk of cardiovascular events.5 A clinical diagnosis of FH is typically made by documenting characteristic clinical features alongside significantly raised LDL-C levels. Additionally, a genetic diagnosis can be confirmed through the detection of heterozygous or biallelic pathogenic variants, primarily in the LDLR, APOB, or PCSK9 genes.6 Nevertheless, AI techniques may have the potential to improve FH identification.

Using only a basic lipid profile, age, and sex, Hesse et al. demonstrated that an ML model more effectively identified genetically confirmed FH in a cohort of individuals suspected of having FH than did LDL-C cutoff values, and it performed comparably to the Dutch Lipid Clinic Network criteria.7 The ML model achieved an area under the receiver operating characteristic curve (AUROC) of 0.711 in an external dataset with a high FH prevalence (n=1,376; FH prevalence=64%), outperforming the LDL-C cutoff (AUROC=0.642) and matching the Dutch Lipid Clinic Network criteria (AUROC=0.705). Furthermore, the model demonstrated higher accuracy when evaluated on individuals with medium or lower FH prevalence, with AUROC values of 0.801 and 0.856, respectively.

Similarly, another study employed three ML algorithms to predict the presence of FH genetic mutations in two independent European cohorts with FH.8 Pina et al.8 discovered that the three ML algorithms outperformed the clinical Dutch Lipid Score in predicting carriers of FH-causative mutations, with AUROC values of 0.79, 0.83, and 0.83 for the Gothenburg cohort, and 0.70, 0.78, and 0.76 for the Milan cohort, compared to the Dutch Lipid Score’s AUROC of 0.68 and 0.64 for the Gothenburg and Milan cohorts, respectively.

Nolde et al.9 demonstrated that a model based on neural networks outperformed traditional clinical diagnostic criteria in predicting pathogenic mutations responsible for FH. The AUROC for the neural network model in predicting an FH gene variant was 0.87, significantly higher than that of other clinical criteria (Dutch Lipid Score: AUROC of 0.743; MEDPED: AUROC of 0.764; Simon Broome: AUROC of 0.635). The model’s global accuracy in predicting mutations causing FH was 80.6%, with a specificity of 83.3%, sensitivity of 74.3%, negative predictive value of 88.5%, and positive predictive value of 66.0%.

Finally, Banda et al.10 developed an ML-based classifier to identify potential FH patients using electronic health record data. The classifier achieved a positive predictive value of 0.88, sensitivity of 0.75, and specificity of 0.99 for detecting genetically confirmed FH cases. In addition, the AUROC was more informative for low-prevalence outcomes.

Underdiagnosis and undertreatment of FH are significant problems affecting the management of the condition. The cardiovascular consequences of FH often remain silent until they manifest as premature cardiovascular mortality and morbidity. Therefore, altering the natural history of FH depends on early detection and the initiation of lipid-lowering therapies. The studies analyzed in this review suggest that an ML-based approach could improve the identification rate of index FH cases. The potential applicability of these novel diagnostic tools for FH could be significant in the future.

In clinical settings, LDL-C is typically estimated using the Friedewald equation. However, this method is known to be inaccurate in cases of high triglycerides, non-fasting states, or when patients exhibit very low LDL-C values.11 To account for variations in triglyceride levels, alternative formulas have been developed to estimate LDL-C more accurately than traditional methods. These include the Martin-Hopkins formula and the Sampson equation.

In this context, several studies have evaluated various AI techniques for estimating LDL-C levels. Oh et al.12 utilized a substantial single-center electronic health record database to develop an ML algorithm that estimates LDL-C from standard lipid profiles, analyzing 823,657 tests. The ML algorithms outperformed the traditional Friedewald and the more recent Martin-Hopkins equations in estimating LDL-C. ML algorithms employing gradient boosting (LDL-CX) and neural networks (LDL-CN) demonstrated a stronger correlation with directly measured LDL-C than the conventional methods, with correlation coefficients of 0.9662 and 0.9668 for LDL-CX and LDL-CN, respectively, compared to 0.9563 and 0.9585 for the Friedewald and Martin-Hopkins equations. Furthermore, the overall bias was significantly lower for LDL-CX (−0.27 mg/dL) and LDL-CN (−0.01 mg/dL) than for the Friedewald formula (−3.80 mg/dL) and Martin-Hopkins equation (−2.00 mg/dL), particularly at elevated triglyceride levels.

A retrospective study was conducted to compare the performance of an ML algorithm with that of a direct homogeneous LDL-C assay.13 The principal findings indicated that the ML algorithm demonstrated better agreement with the direct homogeneous LDL-C assay than other equations, particularly in cases of mild and severe hypertriglyceridemia. Overall, the intraclass correlation coefficients for the estimated LDL-C values compared to the directly measured LDL-C values were 0.894, 0.937, 0.935, 0.869, and 0.925 for the Friedewald formula, the Martin-Hopkins equation, the Sampson equation, the de Cordova equation, and the ML-based estimation, respectively (p<0.001).

Similarly, another study introduced a novel method for estimating LDL-C from the standard lipid profile using an ML approach, which was based on 17,500 lipid profiles from 10,936 individuals.14 The authors found that the correlation coefficients between the estimated and measured LDL-C values were 0.982 for the AI model, surpassing the 0.950 achieved by the Friedewald equation and the 0.962 by the Martin-Hopkins method. Furthermore, the AI model demonstrated superior performance across various subgroups categorized by LDL-C and triglyceride levels. This included individuals with triglycerides greater than 500 mg/dL, where the mean difference in LDL-C estimation was −27.17 mg/dL compared to the Friedewald equation and −4.44 mg/dL compared to the Martin-Hopkins method. Similarly, for individuals with LDL-C levels below 70 mg/dL, the mean difference was −3.82 mg/dL when compared to the Friedewald equation and −1.84 mg/dL compared to the Martin-Hopkins method.

Ghayad et al.15 evaluated an ML algorithm based on age, sex, healthcare setting, and triglyceride levels against a direct LDL-C assay. The ML algorithm demonstrated good agreement with direct LDL-C measurements in patients with normal triglyceride levels and mild hypertriglyceridemia, as indicated by an intraclass correlation coefficient greater than 0.9. However, its performance was less robust in patients with severe hypertriglyceridemia and in those with very low LDL-C levels, where the intraclass correlation coefficient was slightly below 0.9.

Another study applied ML techniques to three databases: health check-up participants at the Resource Center for Health Science (n=2,664), clinical patients at Gifu University Hospital (n=7,409), and clinical patients at Fujita Health University Hospital (n=14,842).16 Subsequently, nine different ML models were developed. A separate test dataset (n=3,711) from Fujita Health University Hospital was used to compare and validate the models against the Friedewald formula and the Martin method. The coefficients of determination for the models using the health check-up dataset were equal to or lower than those of the Martin method. However, the coefficients of determination for several models using the clinical patient data surpassed those of the Martin method. Thus, the study's results indicate that the performance of ML models varies with the data source, which complicates the task of confirming whether new techniques outperform traditional methods.

Finally, Kwon et al.17 aimed to develop a deep neural network model for estimating LDL-C levels and to compare its performance with that of previous LDL-C estimation equations. They used two large, independent datasets from Korean populations, comprising 176,400 individuals. The DL method exhibited lower bias and error than the Friedewald, Martin, and Sampson equations, demonstrating high agreement with LDL-C levels measured by a homogeneous assay.

Considering these reports, new AI-based techniques could offer a valuable approach for estimating LDL-C levels. These methods may not only surpass the traditional Friedewald formula but could also outperform alternatives such as the Martin-Hopkins formula or the Sampson equation. Nonetheless, these models need further refinement and validation across diverse populations. Enhanced LDL-C estimation would lead to more accurate cardiovascular risk stratification and optimize lipid-lowering treatments, ultimately helping to meet the lipid targets suggested by clinical guidelines and reduce patients’ cardiovascular risk.

Based on scientific evidence, current guidelines recommend using LDL-C levels as the main therapeutic goals.18, 19 However, despite advances in lipid-lowering treatments, many patients do not reach the goals recommended by the guidelines.20

In this context, a study evaluated various predictors of lipid goal achievement in outpatients with type 2 diabetes using ML.21 The study conducted a real-world analysis of the lipid profiles of 11,252 patients, employing an ML model to identify the most relevant factors for predicting the achievement of an LDL-C level below 100 mg/dL within two years of initiating lipid-lowering therapy. The model exhibited strong predictive capabilities, with an AUROC of 0.79. The most significant predictors of achieving the treatment goal were baseline LDL-C values, their reduction after 6 months, and no therapy discontinuation. Other factors associated with a higher likelihood of success were elevated high-density lipoprotein cholesterol, presence of albuminuria, body mass index, younger age, male sex, lower blood glucose, hemoglobin A1c levels, and the use of antihypertensive medication. Notably, the ML model also determined the minimum LDL-C reduction required by the 6-month follow-up to improve the chances of meeting the therapeutic target within 2 years. For instance, to achieve the two-year goal, the model predicted that patients with baseline LDL-C levels ranging from 100 to 125 mg/dL, 150 to 175 mg/dL, and over 200 mg/dL would need to reduce their LDL-C by at least 14%, 33%, and 47%, respectively, by the 6-month mark. The authors suggested that these insights could be a valuable asset for guiding therapeutic decisions and promoting more comprehensive analysis and validation. However, they acknowledged limitations such as potential biases, the challenge of generalizing results to different populations, the exclusion of certain data such as statin dosages or some socioeconomic factors, and issues related to non-adherence. These caveats underscore the necessity for additional research, which should also encompass other high-cardiovascular-risk groups, including patients with a history of cardiovascular disease, chronic renal failure, or FH.

Statin intolerance remains an important clinical challenge, although the prevalence of this clinical scenario might often be overestimated.22 Importantly, the nocebo effect could play a key role in this significant public health problem.23

A recent study demonstrated the potential of an AI approach to sift through extensive social media data, providing insights into public perceptions of statins.24 This qualitative research focused on discussions related to statins on a social media platform. A total of 10,233 statin-related discussions authored by 5,188 individuals were identified. The three most prevalent topics within these discussions included concerns about elevated LDL-C levels while following a ketogenic diet, requests for advice and shared experiences regarding changes in lipid panels, and anecdotal views on the effectiveness and side effects of statins. Sentiment analysis conducted on the social media posts and discussions about statins indicated that the overall sentiment was mainly neutral to negative, with 30.8% of posts being negative, 66.6% neutral, and only 2.6% positive.

Sarraju et al.25 developed a DL model to classify statin nonuse and the reasons for it, utilizing unstructured electronic health records from a cohort of 56,530 patients with atherosclerotic cardiovascular disease. The DL model classifiers were able to identify statin nonuse with an AUROC of 0.94 and the reasons for nonuse with a weighted-average AUROC of 0.88, as evaluated against manual expert chart review. This DL-based approach pinpointed key patient-level reasons, such as side effects and patient preference, as well as clinician-level reasons, including guideline-discordant practices, for statin nonuse. It also highlighted differences by patient race or ethnicity. Similarly, a novel DL approach accurately identified statin nonuse in patients with diabetes (n=33,461).26 This method also classified reasons for statin nonuse from unstructured electronic health record data, which included patient reasons (side effects and statin hesitancy), clinician reasons (guideline-discordant practice), and system reasons (clinical inertia), with variations observed by patient race and ethnicity. Older individuals (>75 years of age) were more likely to experience statin-associated side effects or contraindications (23.4%) and were less likely to be affected by clinical inertia or guideline-discordant practice compared to younger individuals (p<0.05 for comparisons). Hispanic patients were the most likely to encounter guideline-discordant practice (24.7%, p<0.05), while Black patients were the most likely to be subject to clinical inertia (24.0%, p<0.05), in comparison to other racial groups.

Interestingly, the last two studies identified some common reasons for statin non-adherence, such as adverse effects or perceived inadequate lipid control/guideline-discordant practice, while other reasons varied, including patient preference, statin hesitancy, or clinical inertia. However, all these reasons have been previously documented in the medical literature.27 Therefore, accurately identifying these barriers using AI techniques could enhance statin utilization among patients with high cardiovascular risk in real-world settings, offering a strategy to bridge significant gaps in dyslipidemia treatment. It is crucial that future research be tailored to the diverse geographical areas and cultural characteristics of each population.

ApoB is an adequate representative of all atherogenic particles.28 When there is a discrepancy between LDL-C and ApoB levels, elevated ApoB has been shown to be a more accurate predictor of atherosclerotic cardiovascular disease risk.29 Webb et al.30 examined the influence of adiposity, diet, and inflammation on the discordance between LDL-C and ApoB by employing an ML model. In this instance, the ML model was applied to data from the National Health and Nutrition Examination Survey to explore cardiometabolic and dietary factors associated with concordance or discordance between LDL-C and ApoB. The ML analysis revealed that body mass index, dietary saturated fatty acids, dietary fiber, serum C-reactive protein, and uric acid were the variables most strongly associated (R²=0.70) with the pattern of low LDL-C and high ApoB. Accurately identifying patients with lipid discordance is crucial, as their condition has been linked to a higher prevalence of atherosclerosis.31, 32

A summary of the main characteristics of the studies analyzed in this review is presented in Table 1.

Table 1
Characteristics of the studies included

• Click for larger image
• Click for full table
• Download as Excel file

Adapting new AI-based techniques to meet patient needs presents a significant challenge. Nevertheless, it is essential to address certain concerns associated with AI techniques. Ethical issues, such as patient privacy, must be taken into consideration.33 Additionally, when interpreting information, one should be mindful of publication-related issues, including a lack of transparency or methodological clarity that can impact the reproducibility of results.34 Familiarizing oneself with the specific methodological aspects of new techniques, such as mathematical models and software, will pose a challenge for the medical community. A key distinction is that conventional statistics are model-driven, whereas AI and ML are data-driven, operating without a predefined understanding of the relationship between data and outcomes.35 In AI and ML, the software identifies patterns and forms clusters of data with shared characteristics that may influence the outcome. Traditional statistical methods, on the other hand, presuppose knowledge of the model that generated the data, assuming a known relationship between input variables and the output. The issue here is that the relationship between input and output is selected by the user and may lead to a less than optimal model. In contrast, ML methods do not start with a presumed model; instead, they begin with the data, and a mathematical algorithm develops a model with prediction as the primary objective. Compared to traditional statistical methods, ML algorithms can manage more variables but also necessitate a larger sample size for analysis. Another challenge is integrating these technologies into existing workflows. AI is not intended to replace clinical judgment. Rather, the most effective use of AI in clinical settings occurs when it is integrated seamlessly into the clinical workflow.36 Finally, it is crucial to consider the specific characteristics of different geographical regions, cultures, and ethnic groups. In this regard, some authors have suggested the creation of a framework to guide research groups in designing, conducting, and reporting their studies; to assist editors and peer reviewers in evaluating contributions to the literature; and to enable patients, clinicians, and policymakers to critically assess where new findings may offer patient benefits.37

The concerns and opportunities related to AI techniques in the area of lipidology are shown in Fig. 1.

Fig. 1
Concerns and opportunities of using AI techniques in the field of lipidology.
AI, artificial intelligence.

• Click for larger image
• Download as PowerPoint slide

In this literature review, we analyzed potential applications of AI, ML, and DL in lipidology. Diagnostic applications, such as screening for FH or calculating LDL-C, could significantly impact clinical practice. Understanding the barriers patients face with lipid-lowering medications could enable physicians to address these obstacles effectively. Furthermore, the development of improved predictive models based on lipid profiles may be anticipated in the near future. Additionally, novel AI techniques could uncover previously unknown disease correlations and facilitate the delivery of precision medicine. However, the social, methodological, and ethical complexities associated with these applications warrant further investigation and regulation.