Artificial intelligence (AI) is revolutionizing medicine by allowing personalized approaches that tailor treatments to individual patients based on their medical history, lifestyle factors, and unique genetic makeup to enhance treatment efficacy, improve patient outcomes, and reduce side effects. This article discusses the applications of AI techniques in personalized medicine.
Image credit: SOMKID THONGDEE/Shutterstock
AI Techniques in Personalized Medicine
AI techniques play a crucial role in personalized medicine by improving the precision of disease treatment and drug administration. AI can accurately and quickly analyze genomic sequences to detect genetic conditions early. Currently, AI-based tools are used to identify genetic variants related to diseases such as cancer, enabling healthcare providers to devise personalized treatment plans for patients.
In medicine, predictive analytics can predict disease progression and clinical trajectories of patients, identify high-risk patients, and estimate patient readmission rates. The insights derived from these predictions enable proactive and personalized healthcare delivery.
Moreover, AI techniques can also offer personalized health recommendations based on environmental factors and individual health records. Personalized AI-powered chatbots can remind patients to take medication, schedule appointments, and provide real-time health advice, making patient engagement more efficient and tailored to an individual’s requirements.
Detection/prediction of disease, achievement of accurate diagnosis, and treatment optimization are the major challenges in personalized medicine, which can be addressed using several AI techniques, including Naïve Bayes (NB), artificial neural network (ANN), support vector machine (SVM), and fuzzy logic.
SVM: SVM can be used to detect and classify fall types with 99% accuracy, which is crucial to prevent falling. The technique can also be utilized to detect cardiac monitoring device accuracy. The generic cardiac arrest monitoring systems cannot differentiate between true events and artifacts across different individuals as they have been trained only on a general population level.
SVM can be used to realize a more accurate patient-specific result using the prior population-level knowledge for initial model creation and then refining the initial model by interacting with human experts selectively to obtain examples from a new patient until a stopping condition is fulfilled.
Heart disease can be diagnosed with higher accuracy using SVM, with studies demonstrating that 86.42% accuracy can be attained using SVM compared to only 80.81% accuracy attained using radial basis function (RBF). A hybrid machine learning (ML) methodology based on SVM classifier and bag-of-words representation can be employed to identify and disseminate healthcare information.
ANN: Decision trees (DTs) and neural networks (NNs) can be utilized to develop real-time patient-specific alarm algorithms. However, studies displayed that NN trained for eight hours had a marginally higher accuracy rate than DT and outperformed the generic alarm algorithms in devices. By analyzing spectral information and diagnostic criteria, ANN can also be used to accurately diagnose several diseases, including eye problems and different forms of cancer such as malignant melanoma.
In a recent study, ANN was used to effectively diagnose kidney stones. A dataset consisting of 1000 patient records and seven attributes representing the actual symptoms of kidney stones was used to perform the study.
Additionally, three NN algorithms, including multilayer perceptron (MLP) with backpropagation, learning vector quantization (LVQ), and RBF, were compared. Results demonstrated that MLP outperformed the other algorithms with 92% accuracy by correctly classifying 922 instances and incorrectly classifying 23 instances.
Multiple studies have been performed to predict the Graves disease (GD) clinical course in patients treated using antithyroid drugs. In a study, researchers investigated the clinical outcome of GD after antithyroid therapy by defining a set of easily achievable variables. Perceptron-like ANN was utilized to predict relapse/remission after Methimazole withdrawal. 27 variables obtained during treatment or at diagnosis were considered.
Among various combinations, an optimal set of seven variables available at diagnosis, including cigarette smoking, thyroid-ultrasonography findings, serum TGAb and fT4 levels at presentation, psychological symptoms requiring psychotropic drugs, presence of thyroid bruits, and heart rate, was identified as these variables were crucial for efficiently predicting the disease outcome following antithyroid therapy withdrawal. The results demonstrated that perceptron-like ANN can be an effective tool in GD management for selecting the most appropriate therapy schedule during diagnosis.
Fuzzy Logic: Case-based fuzzy cognitive map (CBFCM) is an extension of fuzzy cognitive maps used for prediction and classification. CBFCM is used in personalized medicine to analyze the relationship between various diseases and patient-specific information, including blood type, blood pressure, and gene, to infer a pattern match in the patient's disease detection approach.
Fuzzy logic has also been utilized to detect heart disease using six input fields, including old peak, maximum heart rate, blood sugar, cholesterol, blood pressure, and chest pain, and two output fields, including precautions and results, with a rule base consisting of 22 rules. The presence of heart disease was detected with 92% accuracy.
Expert Systems: In a study, a rule-based expert system was developed to produce relevant information and data for results, consultations, and possible diagnoses. The study was primarily focused on Alzheimer‟s disease, bronchitis, migraine, typhoid, thyroid, hepatitis, jaundice, asthma, cholera, diabetes, diarrhea, chicken pox, and malaria.
In another study, a rule-based expert system was developed to make inferences using symbols that require translating specific knowledge in standard symbolic form. Data that associate patients with signs/symptoms and diseases was collected and the symptoms were then organized in groups.
Major Applications of AI
Histological features of cancerous tissue primarily diagnose breast cancer prognostics. However, the breast cancer image grading does not include key stromal features for its scoring, which indicates the subjectivity of the grading method.
Using an ML model, automatic classification can detect novel overlooked features and eliminate image reading subjectivity. An automated algorithm, designated as C-path, was developed in a study to detect novel features in the stromal and epithelial part of the image.
In the first step of the algorithm, supervised learning was used to classify images as epithelial or stroma, and then, rich feature sets were obtained from both stromal and epithelial parts. Eventually, images labeled as deceased after five years and survived after five years were utilized to build a 5-Year-Survival (5YS) predictive model. The ML algorithm effectively identified three novel stromal features that can predict 5YS with good accuracy and eight epithelial features.
Although different cancer types possess similar capabilities, such as cell proliferation and invasiveness, identifying the common genes for the cancers has been extremely difficult. In a study, an unsupervised clustering algorithm was developed to identify attractor metagenes using Molecular Taxonomy of Breast Cancer International Consortium data of six cancers with 11,395 total genes.
The attractors identified using this algorithm, including lymphocyte-specific attractors, mitotic chromosomal instability attractors, and mesenchymal transition attractors, were strongly related to three major cancer processes. Heart Failure with preserved Ejection Fraction (HFpEF) is a cardiac disorder with a heterogeneous phenotype. An unsupervised ML algorithm was developed in a study to map HFpEF patients into distinct phenotype groups/phenomapping. Initially, Bayesian information criterion analysis was utilized to identify the optimal number of distinct phenotype groups.
Then, the algorithm was used to sort detailed echocardiographic, electrocardiographic, laboratory, and clinical data from 397 participants into three distinct phenotype groups. Results showed that different phenogroups have different risk factors. The stratification of patients into different phenotype groups can play a crucial role in personalized therapies.
An ML model was built in another study that can predict the antidepressant treatment outcome in patients with major depressive disorder (MDD) using a combination of sociodemographic, genomics, and metabolomics factors. The model was developed using data from the Mayo Clinic Pharmacogenomics Research Network Antidepressant Medical Pharmacogenomic Study with 603 participants/patients who completed the trial. Sociodemographic factors, clinical measures, and biological data were used as predictor variables.
All metabolites identified were associated with mood in behavioral sciences. Overall, the model including sociodemographic factors and omics increased the accuracy of the antidepressant treatment outcome prediction from 35% to 80%, which indicated a substantial improvement compared to the accuracy achievable using only clinical measures.
References and Further Reading
Gifari, M. W., Samodro, P., Kurniawan, D. W. (2021). Artificial Intelligence toward Personalized Medicine. Pharmaceutical Sciences and Research, 8(2), 65 -72. https://scholarhub.ui.ac.id/psr/vol8/iss2/1/
Awwalu, J., Garba, A. G., Ghazvini, A., Atuah, R. (2015). Artificial Intelligence in Personalized Medicine Application of AI Algorithms in Solving Personalized Medicine Problems. International Journal of Computer Theory and Engineering, 7, 439-443. https://doi.org/10.7763/IJCTE.2015.V7.999.
Schaar, M. (2023). AI-powered personalised medicine could revolutionise healthcare. [Online] (Accessed on 27 November 2023)
The Role of Artificial Intelligence in Personalized Medicine: Exploring AI-driven Healthcare Solutions. [Online] (Accessed on 27 November 2023)