Machine learning applications in healthcare clinical practice and research

Abstract

Machine learning (ML) is a type of artificial intelligence that assists computers in the acquisition of knowledge through data analysis, thus creating machines that can complete tasks otherwise requiring human intelligence. Among its various applications, it has proven groundbreaking in healthcare as well, both in clinical practice and research. In this editorial, we succinctly introduce ML applications and present a study, featured in the latest issue of the World Journal of Clinical Cases. The authors of this study conducted an analysis using both multiple linear regression (MLR) and ML methods to investigate the significant factors that may impact the estimated glomerular filtration rate in healthy women with and without non-alcoholic fatty liver disease (NAFLD). Their results implicated age as the most important determining factor in both groups, followed by lactic dehydrogenase, uric acid, forced expiratory volume in one second, and albumin. In addition, for the NAFLD- group, the 5^th and 6^th most important impact factors were thyroid-stimulating hormone and systolic blood pressure, as compared to plasma calcium and body fat for the NAFLD+ group. However, the study's distinctive contribution lies in its adoption of ML methodologies, showcasing their superiority over traditional statistical approaches (herein MLR), thereby highlighting the potential of ML to represent an invaluable advanced adjunct tool in clinical practice and research.

Core Tip: Across numerous diverse industries, machine learning (ML) is revolutionizing healthcare as well. It has demonstrated the potential to aid in disease diagnosis, treatment planning, decision-making, and outcome prediction, as well as improve clinical trial design and their success rates, often surpassing traditional methods. We highlight a study, published in the World Journal of Clinical Cases, where ML techniques proved superior to traditional statistical methods in analyzing factors affecting the estimated glomerular filtration rate in healthy women with and without non-alcoholic fatty liver disease.

INTRODUCTION

Machine learning (ML) is a type of artificial intelligence that assists computers in the acquisition of knowledge through data analysis, thus creating machines that can complete tasks otherwise requiring human intelligence. It encompasses various techniques, among which are neural networks, decision trees, support vector machines, and ensemble algorithms, to name a few[1]. ML allows the generation of predictions or assessments without the need for explicit programming[2]. By using algorithms to analyze and find patterns in large datasets, it enables the development of models that can make accurate predictions or identify insights[3]. These models are trained using supervised (labelled data) or unsupervised (unlabeled data) learning, a hybrid form of the two using both labelled and unlabeled data (semi-supervised learning), and reinforcement learning (based on reward or penalty)[2], and they can improve over time with additional data and feedback.

Unsurprisingly, considering its immense, endless capabilities, ML has demonstrated revolutionary applications across numerous diverse sectors, from robotics[4] to finance[5,6], and from business[7] to cybersecurity[8] and healthcare[9-11]. Some of the largest applications of ML in healthcare include the analysis of medical images[12], natural language processing in electronic health records[13], and using human genetics in disease prediction and disease etiology identification[14].

ML ALGORITHMS

ML algorithms can assist healthcare professionals in diagnosing diseases, developing personalized treatment plans, aiding in decision-making[15], and predicting patient outcomes. ML models such as convolutional neural networks (CNNs) or recurrent neural networks (RNNs) are an important part of ML's significant role in healthcare. Both CNNs and RNNs automate feature extraction and pattern recognition, enabling faster, more accurate diagnoses and predictions compared to traditional methods, thus transforming clinical decision-making. CNNs use layers of filters to automatically learn and extract important features from medical images. Convolutions capture spatial hierarchies, allowing CNNs to identify edges, textures, and patterns indicative of medical conditions. Based on these patterns learnt from vast amounts of training data, eventually CNNs become capable of differentiating between healthy tissue and abnormalities, therefore assisting in tasks like detecting tumors or classifying diseases. On the other hand, RNNs excel in predictive analytics by analyzing sequential data. RNNs maintain memory of previous inputs, making them ideal for predicting disease progression or patient outcomes. In healthcare, by analyzing trends in patient data (i.e., vital signs and laboratory results), RNNs may be used to predict events (i.e., hospital readmission).

Recent developments have shown that the utilization of big data and ML has the potential to generate algorithms demonstrating comparable performance to that of human physicians[16], including tasks such as determining the presence of tumors on radiological and/or other kinds of medical imaging[15,17]. Some examples of applications in medicine that have even gained United States Food and Drug Administration approval include different types of software that may detect cardiac rhythm disorders or heart failure or rapidly and accurately assess radiological images such as X-rays and computed tomography scans for the identification of several conditions, including strokes, fractures, tumors, intracranial hemorrhage, aortic dissection, pneumothorax, suspected uncontained intra-abdominal gas, and lung nodules[18].

To further illustrate their capabilities, ML methods have even proven efficient with tasks such as the utilization of magnetic resonance imaging scans to assist with preoperative tumor staging[19], thereby demonstrating their potential to break the barrier of disease diagnosis and influence advanced decision-making by enabling more accurate treatment options as well (i.e., surgical planning and timing of chemoradiation)[20]. On a similar note, ML methods are often applied to analyze and interpret complex data obtained through hyperspectral imaging, an advanced imaging technique that captures and processes information from multiple spectral bands across the electromagnetic spectrum. In healthcare and medical diagnostics, this advanced technique can provide detailed information that can improve disease diagnosis and treatment by analyzing tissue characteristics non-invasively. For example, ML methods combined with hyperspectral imaging have been successfully employed in detecting skin cancer and in the early detection of esophageal cancer[21,22]. In addition, the ability of ML algorithms to indirectly extrapolate information from previously available datasets has proven capable of allowing opportunistic screening and early diagnosis of diseases, thus allowing for optimized patient outcomes, and synchronously opening new horizons into research opportunities[23,24]. Similarly, within the realm of research, ML has been employed to scrutinize extensive datasets and spot intricate patterns that would otherwise pose challenges for human interpretation, thus leading to significant advancements in various fields (i.e., genetics and pharmaceutical exploration)[15].

On a similar note, ML has exhibited the capacity to augment the design of clinical trials and enhance their success rate through diverse means, such as employing predictive models to be able to uphold optimal statistical power with smaller sample sizes[25,26]. To elaborate on the role of ML in clinical trials, statistics, and research, it is essential to recognize that while ML models should not be considered a panacea compared to traditional statistical methods[27], they are increasingly augmenting or replacing traditional approaches in classifying and predicting health outcomes[28]. For example, ML methods have been used in comparison with traditional statistical methods, where they seemed to outperform conventional logistic regression models in the prediction of abnormal carotid intima-media thickness in patients with type 2 diabetes[29].

To further illustrate the impact of ML advancements in structuring clinical studies and inferring insights that may prove useful for clinical practice while also exploring ML methods' abilities to outperform traditional statistical methods, we discuss the recent study by Chen et al[30], featured in the latest issue of the World Journal of Clinical Cases. The authors of this study conducted an analysis, using both multiple linear regression (MLR) and ML methods, to investigate the significant factors that may impact the estimated glomerular filtration rate (eGFR) in healthy women with and without non-alcoholic fatty liver disease (NAFLD).

NAFLD is a prevalent etiology of chronic liver disease on a global scale and encompasses a range of conditions that are distinguished by the presence of hepatic steatosis, with no other identifiable factors (such as excessive alcohol consumption) contributing to the buildup of fat in the liver. NAFLD may range from the benign non-alcoholic fatty liver to the more severe non-alcoholic steatohepatitis, while it even holds the potential to advance to fibrosis, cirrhosis, and even neoplastic conditions (namely hepatocellular carcinoma)[31]. Evidently, NAFLD carries significant health, economic, and social repercussions[32] and therefore seems to be a significant health concern warranting increasing attention in healthcare. The correlation between NAFLD and various other disorders has received great interest in recent years, with chronic kidney disease (CKD) emerging as a significant connection, both in terms of prevalence and significance[33,34]. Notably, around one-third of NAFLD patients exhibit renal function impairment[35]. The eGFR serves as a valuable and direct indicator of renal filtration function, frequently utilized in clinical settings for diagnosing CKD and evaluating renal function[36]. The eGFR is closely linked to various risk factors [including hypertension, obesity, liver enzymes, lipid levels, uric acid (UA), and hemoglobin] associated with NAFLD in apparently healthy populations. Furthermore, the influence of NAFLD on eGFR may be influenced by age, a significant risk factor for kidney disease, given that eGFR calculation incorporates age as one of its key parameters[37]. However, the majority of studies focusing on the above matters have relied on traditional statistical analysis methods, with limited research utilizing ML approaches. Only recently, Cao et al[38] employed ML to identify NAFLD-related genes as diagnostic markers in CKD patients with NAFLD, highlighting the potential of ML in this context.

Thus, Chen et al's recent study[30] represents a significant effort to address the research gap in the use of ML techniques to investigate risk factors affecting eGFR in NAFLD patients. Specifically, the authors of this retrospective study employed a combination of traditional statistical methods (MLR) and three ML techniques (stochastic gradient boosting, extreme gradient boosting, and elastic net) to analyze a large and diverse dataset covering over 100 biological indicators and identify key factors influencing eGFR in healthy Chinese women, both with and without NAFLD (NAFLD+ and NAFLD-). The study found that ML methods surpassed MLR in performance. Age emerged as the most crucial factor influencing eGFR in both NAFLD- and NAFLD+ groups, followed by lactic dehydrogenase (LDH), UA, forced expiratory volume in one second (FEV1), and albumin (Alb). In the NAFLD- group, thyroid-stimulating hormone (TSH) and systolic blood pressure (SBP) were ranked as the 5^th and 6^th most significant factors, whereas plasma calcium (Ca) and body fat (BF) were identified as the 5^th and 6^th key factors in the NAFLD+ group.

The study's findings align well with existing literature while simultaneously building upon it. It's unsurprising that age emerged as the primary determinant for eGFR, given the well-documented decline in renal function with advancing age[37]. Notably, elevated LDH levels have been identified as a predictive marker for renal failure[39], and elevated UA levels can compromise renal function, potentially leading to CKD through various pathways[40]. Additionally, FEV1, a useful indicator of pulmonary function, also appears relevant, as obstructive lung function has been associated with an increased risk of renal impairment[41]. Alb levels have also been consistently linked with renal function across several studies[42]. TSH has been shown to influence renal function, warranting regular monitoring in patients with hypothyroidism[43] and demonstrating improvement with appropriate treatment[44]. Furthermore, increased SBP is a well-established independent risk factor for CKD and end-stage renal disease[45,46]. Finally, low serum Ca levels have been implicated as an independent factor for CKD progression[47] and linked to poor renal outcomes[48], while BF has also been shown to negatively impact renal function, aiding in CKD progression, and has been recommended as a surveillance marker for renal impairment[49].

The study’s findings have noteworthy implications for understanding the relationship between NAFLD and renal function and shed light on the importance of the various factors found to significantly affect eGFR in NAFLD+ and NAFLD- patients. Considering that the study participants represent a healthy cohort, these insights could potentially facilitate early detection and management of renal diseases in at-risk populations, thereby playing a crucial role in population screening. The ML approach employed in the study seems to have improved the robustness and reliability of the results. By utilizing three distinct ML methods that provided valuable insights and surpassed the traditional MLR model in, the study highlights the potential of advanced analytics in medical research.

As acknowledged by the authors, the study presents certain limitations, including its exclusive focus on healthy Chinese women and its cross-sectional design. To enhance the study's broader applicability and generalizability, conducting longitudinal research and expanding the methodology to encompass both men and women from diverse ethnic backgrounds would be desirable. Furthermore, expanding the research beyond healthy populations to explore how various factors influence eGFR across different clinical conditions and diseases could yield invaluable insights. Additionally, given that grades 1 to 3 were collectively defined as fatty liver based on ultrasound findings in this study, conducting subgroup analyses to explore the factors influencing eGFR across different grades of fatty liver could provide interesting results as well.

CONCLUSION

Overall, the current study presents a useful contribution to the understanding of factors affecting eGFR in healthy women with and without NAFLD. The use of ML techniques and the demonstration of its superiority compared to traditional statistics add the most novel dimension to the present study, potentially paving the way for ML to become an invaluable advanced adjunct tool in clinical practice and research. Nonetheless, despite its promising potential, integrating ML into clinical practice and research requires awareness and caution as it also presents challenges, including the necessity for robust validation, interpretability of models, and data privacy concerns. To provide additional context, robust validation is necessary because ML models must be tested carefully on diverse, high-quality datasets to confirm their accuracy, generalizability, and reliability in real-world medical scenarios. Without proper validation, inaccurate predictions could lead to erroneous decisions. Interpretability is another key concern. The decision-making processes of many ML models, especially DL models, are difficult for clinicians to understand and therefore adopt their results in clinical practice. To build trust and enable acceptance in clinical practice, ML models must provide clear, explainable outputs that physicians can interpret and justify. Finally, data privacy is critical. Medical datasets contain sensitive patient information, and integrating ML requires secure handling, adhering to appropriate regulations. Ensuring data security while maintaining access to large, comprehensive datasets for training models is a delicate balance that needs careful management.