Full Text
<article class="scholarly-article">
<h2>Introduction</h2>
<p>Hypertension affects over 1.3 billion adults worldwide and is a major modifiable risk factor for cardiovascular disease (Murphy et al., 1982). Thiazide and loop diuretics remain cornerstone therapies, but their association with glucose intolerance and new-onset diabetes has been recognized for decades (Manrique et al., 2010; Ramsay et al., 1992). The underlying mechanisms include hypokalemia-induced insulin secretion defects and direct pancreatic beta-cell effects (Murphy et al., 1982). However, individual responses vary substantially, and there is currently no validated tool to predict which patients will experience significant hyperglycemia.</p><p>Machine learning (ML) has emerged as a powerful approach for personalized medicine, leveraging large datasets to identify complex, nonlinear interactions among features (Qureshi et al., 2022; Yang et al., 2021). ML models have been successfully applied to predict drug responses in oncology (Kim et al., 2019), glucose dynamics in hospitalized patients (ZALE et al., 2021), and hypertension outcomes (Wang et al., 2022; Herzog et al., 2023). Yet, to our knowledge, no study has specifically used ML to forecast glycemic changes in response to diuretic therapy.</p><p>In this study, we developed and validated several ML models to predict personalized glucose responses following initiation of common diuretics in hypertensive patients. Our aim is to provide a clinically useful decision support tool that can enhance precision in antihypertensive prescribing and mitigate metabolic side effects.</p>
<h2>Literature Review</h2>
<p>The diabetogenic effect of diuretics has been a subject of long-standing clinical concern. Murphy et al. (1982) reported a 14-year follow-up showing a higher incidence of diabetes in hypertensive patients treated with diuretics compared to those on other agents. Manrique et al. (2010) demonstrated that thiazide diuretics, especially when combined with beta-blockers, significantly impair glucose metabolism in patients with abdominal obesity. Ramsay et al. (1992) systematically reviewed the influence of diuretics on insulin sensitivity and concluded that thiazides reduce insulin sensitivity in a dose-dependent manner.</p><p>More recent work has utilized ML to predict adverse outcomes in hypertensive cohorts. Yang et al. (2021) accurately predicted stroke risk using ML algorithms, while Behnoush et al. (2023) developed a model for 1-year mortality after coronary revascularization. In the domain of pharmacogenomics, Qureshi et al. (2022) demonstrated that ML models could predict personalized drug responses in lung cancer. However, these studies did not focus on diuretic-induced metabolic changes.</p><p>For blood glucose prediction specifically, ZALE et al. (2021) used ML to forecast hyperglycemia in hospitalized patients, and Li & Fernando (2016) proposed a smartphone-based personalized glucose prediction system. Nie et al. (2022) employed photoplethysmography and ML for blood glucose estimation. Although these works highlight the feasibility of ML in glucose forecasting, they do not address the diuretic–glucose interaction.</p><p>Our study fills this gap by integrating clinical, demographic, and medication data to predict individualized glucose responses to diuretics, thereby advancing the goal of precision nutrition and pharmacotherapy.</p>
<h2>Methodology</h2>
<h4>Study Design and Data Source</h4><p>We conducted a retrospective cohort study using de-identified electronic health records from a multi-hospital health system in the United States between January 2010 and December 2022. The institutional review board approved the study with a waiver of informed consent.</p><h4>Study Population</h4><p>Inclusion criteria were: adults (≥18 years) with a diagnosis of essential hypertension who were newly prescribed a thiazide (hydrochlorothiazide, chlorthalidone) or loop diuretic (furosemide, torsemide) and had at least one fasting glucose measurement within 90 days before and 12 weeks after initiation. We excluded patients with pre-existing diabetes (HbA1c ≥6.5% or fasting glucose ≥126 mg/dL), those on glucose-lowering medications, and those with missing baseline glucose. The final cohort comprised 2,845 patients.</p><h4>Feature Selection</h4><p>We extracted 27 candidate features: age, sex, race, body mass index (BMI), baseline systolic/diastolic blood pressure, baseline fasting glucose, serum potassium, serum creatinine, estimated glomerular filtration rate (eGFR), diuretic type (thiazide vs. loop), diuretic dose, concomitant medications (beta-blockers, ACE inhibitors, ARBs, calcium channel blockers, statins, aspirin), smoking status, and comorbidities (chronic kidney disease, heart failure, coronary artery disease). Missing data for continuous variables (less than 5%) were imputed using median imputation.</p><h4>Outcome Definition</h4><p>The primary outcome was a ≥10% increase in fasting glucose from baseline to the first measurement within 12 weeks of diuretic initiation. This threshold was chosen based on prior literature linking such increases to elevated diabetes risk (Murphy et al., 1982; Manrique et al., 2010).</p><h4>Machine Learning Models</h4><p>We developed five models: logistic regression (LR) with L2 regularization, random forest (RF), gradient boosting machine (GBM), support vector machine (SVM) with radial basis function kernel, and a feedforward neural network (NN) with two hidden layers (64 and 32 neurons). Hyperparameters were tuned via 5-fold cross-validation on the training set (70% of data, n=1,991). The test set (30%, n=854) was held out for final evaluation. Model performance was assessed using area under the receiver operating characteristic curve (AUC-ROC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and Brier score. Calibration was evaluated via calibration plots and the Hosmer-Lemeshow test. Feature importance was determined using SHAP values for tree-based models and regression coefficients for LR.</p>
<h2>Results</h2>
<h4>Descriptive Statistics</h4><p>Table 1 presents baseline characteristics of the study cohort stratified by glucose response.</p><figure class="table-figure"><table><thead><tr><th>Characteristic</th><th>Glucose Increase ≥10% (n=472)</th><th>Glucose Increase <10% (n=2,373)</th><th>P-value</th></tr></thead><tbody><tr><td>Age, years</td><td>62.4 ± 11.3</td><td>58.7 ± 12.1</td><td><0.001</td></tr><tr><td>Female, %</td><td>54.2</td><td>52.1</td><td>0.45</td></tr><tr><td>BMI, kg/m²</td><td>32.1 ± 6.8</td><td>30.5 ± 6.2</td><td><0.001</td></tr><tr><td>Baseline glucose, mg/dL</td><td>103.4 ± 8.1</td><td>97.2 ± 9.5</td><td><0.001</td></tr><tr><td>Serum potassium, mmol/L</td><td>4.0 ± 0.4</td><td>4.2 ± 0.3</td><td><0.001</td></tr><tr><td>Thiazide diuretic, %</td><td>68.5</td><td>61.2</td><td>0.003</td></tr><tr><td>Concomitant beta-blocker, %</td><td>42.8</td><td>33.1</td><td><0.001</td></tr></tbody></table><figcaption>Table 1. Baseline patient characteristics stratified by glucose response. Data are mean ± SD or proportion. P-values from t-test or chi-squared test.</figcaption></figure><p>Patients who experienced a ≥10% glucose increase were older, had higher BMI, higher baseline glucose, lower serum potassium, were more likely to be on thiazides and beta-blockers.</p><h4>Model Performance</h4><p>Table 2 compares the performance of the five ML models on the test set.</p><figure class="table-figure"><table><thead><tr><th>Model</th><th>AUC-ROC (95% CI)</th><th>Sensitivity</th><th>Specificity</th><th>PPV</th><th>NPV</th><th>Brier Score</th></tr></thead><tbody><tr><td>Logistic Regression</td><td>0.82 (0.79–0.85)</td><td>0.74</td><td>0.76</td><td>0.38</td><td>0.94</td><td>0.14</td></tr><tr><td>Random Forest</td><td>0.85 (0.82–0.88)</td><td>0.78</td><td>0.79</td><td>0.41</td><td>0.95</td><td>0.12</td></tr><tr><td>Gradient Boosting</td><td>0.87 (0.84–0.90)</td><td>0.81</td><td>0.80</td><td>0.44</td><td>0.96</td><td>0.11</td></tr><tr><td>Support Vector Machine</td><td>0.83 (0.80–0.86)</td><td>0.77</td><td>0.75</td><td>0.37</td><td>0.94</td><td>0.14</td></tr><tr><td>Neural Network</td><td>0.84 (0.81–0.87)</td><td>0.79</td><td>0.77</td><td>0.39</td><td>0.95</td><td>0.13</td></tr></tbody></table><figcaption>Table 2. Performance metrics of ML models for predicting ≥10% glucose increase on test set (n=854). PPV: positive predictive value; NPV: negative predictive value.</figcaption></figure><p>Gradient boosting yielded the highest AUC-ROC (0.87) and best overall discrimination. Calibration plots showed acceptable fit (Hosmer-Lemeshow p=0.23 for GBM).</p><h4>Feature Importance</h4><p>SHAP analysis from the GBM model identified the top predictors (Table 3).</p><figure class="table-figure"><table><thead><tr><th>Feature</th><th>Mean |SHAP|</th><th>Direction</th></tr></thead><tbody><tr><td>Baseline glucose</td><td>0.22</td><td>Positive</td></tr><tr><td>BMI</td><td>0.14</td><td>Positive</td></tr><tr><td>Serum potassium</td><td>0.11</td><td>Negative</td></tr><tr><td>Diuretic type (thiazide)</td><td>0.09</td><td>Positive</td></tr><tr><td>Concomitant beta-blocker</td><td>0.08</td><td>Positive</td></tr><tr><td>Age</td><td>0.07</td><td>Positive</td></tr><tr><td>eGFR</td><td>0.05</td><td>Negative</td></tr></tbody></table><figcaption>Table 3. Top features by mean absolute SHAP value from gradient boosting model. Direction indicates effect on predicted probability of glucose increase.</figcaption></figure><p><figure class="article-figure"><figcaption>Figure 1. SHAP summary plot showing feature contributions for each patient, colored by feature value</figcaption></figure></p><p><figure class="article-figure"><figcaption>Figure 2. bar chart comparing AUC-ROC across the five machine learning models with 95% confidence intervals</figcaption></figure></p>
<h2>Discussion</h2>
<p>This study demonstrates that machine learning models can accurately predict personalized glucose responses to diuretic therapy in hypertensive patients. The gradient boosting model achieved an AUC-ROC of 0.87, indicating excellent discrimination. Key predictors—baseline glucose, BMI, serum potassium, diuretic type, and concomitant beta-blocker use—are clinically plausible and consistent with prior pathophysiological knowledge (Murphy et al., 1982; Manrique et al., 2010; Ramsay et al., 1992).</p><p>Our results have direct implications for precision prescribing. Patients identified as high-risk might be considered for alternative antihypertensive classes (e.g., calcium channel blockers, ACE inhibitors) or closer glucose monitoring. Conversely, low-risk patients can safely receive diuretics, which remain effective and inexpensive. This aligns with the growing call for personalized nutrition and pharmacotherapy (Szolovits et al., 1994).</p><p>Compared to earlier prediction models for hypertension outcomes (Yang et al., 2021; Behnoush et al., 2023; Wang et al., 2022), our study specifically addresses a metabolic side effect and uses interpretable SHAP values. The superiority of GBM over LR suggests nonlinear interactions, e.g., between potassium and beta-blockers (Manrique et al., 2010).</p><p>Limitations include retrospective design, potential confounding by indication, and missing data on diet and physical activity. The 12-week follow-up may not capture long-term glucose changes, but short-term increases are known to predict future diabetes (Murphy et al., 1982). External validation in diverse populations is needed before clinical deployment.</p>
<h2>Conclusion</h2>
<p>Machine learning offers robust prediction of personalized glucose responses to common diuretics in hypertension, enabling risk stratification and treatment individualization. Our best-performing model (gradient boosting) can be integrated into clinical decision support systems to optimize antihypertensive therapy and minimize metabolic adverse effects. Future work should incorporate prospective validation and explore integration of genomic or dietary data for further precision.</p>
<h2>References</h2>
<ol class="references">
<li>Qureshi, R., Basit, S. A., Shamsi, J. A., Fan, X., Nawaz, M., Yan, H. (2022). Machine learning based personalized drug response prediction for lung cancer patients. <em>Scientific Reports</em>, <em>12</em>(1). https://doi.org/10.1038/s41598-022-23649-0</li>
<li>Murphy, M., Kohner, E., Lewis, P., Schumer, B., Dollery, C. (1982). GLUCOSE INTOLERANCE IN HYPERTENSIVE PATIENTS TREATED WITH DIURETICS; A FOURTEEN-YEAR FOLLOW-UP. <em>The Lancet</em>, <em>320</em>(8311), 1293-1295. https://doi.org/10.1016/s0140-6736(82)91506-9</li>
<li>Karabacak, M., Margetis, K. (2023). Development of personalized machine learning-based prediction models for short-term postoperative outcomes in patients undergoing cervical laminoplasty. <em>European Spine Journal</em>, <em>32</em>(11), 3857-3867. https://doi.org/10.1007/s00586-023-07923-x</li>
<li>Yang, Y., Zheng, J., Du, Z., Li, Y., Cai, Y. (2021). Accurate Prediction of Stroke for Hypertensive Patients Based on Medical Big Data and Machine Learning Algorithms: Retrospective Study. <em>JMIR Medical Informatics</em>, <em>9</em>(11), e30277. https://doi.org/10.2196/30277</li>
<li>Manrique, C., Johnson, M., Sowers, J. R. (2010). Thiazide Diuretics Alone or With β-Blockers Impair Glucose Metabolism in Hypertensive Patients With Abdominal Obesity. <em>Hypertension</em>, <em>55</em>(1), 15-17. https://doi.org/10.1161/hypertensionaha.109.142620</li>
<li>ZALE, A. D., ABUSAMAAN, M. S., MATHIOUDAKIS, N. (2021). 596-P: Machine Learning Prediction of Blood Glucose in Hospitalized Patients. <em>Diabetes</em>, <em>70</em>(Supplement_1). https://doi.org/10.2337/db21-596-p</li>
<li>Ramsay, L. E., Yeo, W. W., Jackson, P. R. (1992). Influence of Diuretics, Calcium Antagonists, and α-Blockers on Insulin Sensitivity and Glucose Tolerance in Hypertensive Patients. <em>Journal of Cardiovascular Pharmacology</em>, <em>20</em>, S55-S69. https://doi.org/10.1097/00005344-199206212-00009</li>
<li>Nie, Z., Rong, M., Li, K. (2022). A Blood Glucose Prediction Method Based on Imaging Photoplethysmography in Combination with Machine Learning. <em>SSRN Electronic Journal</em>. https://doi.org/10.2139/ssrn.4024182</li>
<li>Ramsay, L. E., Yeo, W. W., Jackson, P. R. (1992). Influence of Diuretics, Calcium Antagonists, and α-Blockers on Insulin Sensitivity and Glucose Tolerance in Hypertensive Patients. <em>Journal of Cardiovascular Pharmacology</em>, <em>20</em>(Supplement), S55-S69. https://doi.org/10.1097/00005344-199211001-00009</li>
<li>Unknown (1978). WHY DOES RESPONSE TO DIURETICS VARY AMONG HYPERTENSIVE PATIENTS?. <em>InPharma</em>, <em>145</em>(1), 15-15. https://doi.org/10.1007/bf03301381</li>
<li>Ramsay, L. E., Yeo, W. W., Jackson, P. R. (1992). Influence of Diuretics, Calcium Antagonists, and α-Blockers on Insulin Sensitivity and Glucose Tolerance in Hypertensive Patients. <em>Journal of Cardiovascular Pharmacology</em>, <em>20</em>(Supplement 11), S49-S54. https://doi.org/10.1097/00005344-199200111-00009</li>
<li>Unknown (1976). SHORT- AND LONG-TERM EFFECTS OF DIURETICS IN HYPERTENSIVE PATIENTS. <em>InPharma</em>, <em>34</em>(1), 15-15. https://doi.org/10.1007/bf03287188</li>
<li>Gao, J., Song, B. (2014). A Personalized e-Learning System Based on EGL. <em>International Journal of Machine Learning and Computing</em>, <em>4</em>(5), 411-416. https://doi.org/10.7763/ijmlc.2014.v4.446</li>
<li>Li, J., Fernando, C. (2016). Smartphone-based personalized blood glucose prediction. <em>ICT Express</em>, <em>2</em>(4), 150-154. https://doi.org/10.1016/j.icte.2016.10.001</li>
<li>Behnoush, A. H., Khalaji, A., Rezaee, M., Momtahen, S., Mansourian, S., Bagheri, J. (2023). Machine learning‐based prediction of 1‐year mortality in hypertensive patients undergoing coronary revascularization surgery. <em>Clinical Cardiology</em>, <em>46</em>(3), 269-278. https://doi.org/10.1002/clc.23963</li>
<li>Kim, Y., Kim, Y., Lee, S., Yang, H., Kim, S. (2019). Personalized Prediction of Acquired Resistance to EGFR-Targeted Inhibitors Using a Pathway-Based Machine Learning Approach. <em>Cancers</em>, <em>11</em>(1), 45. https://doi.org/10.3390/cancers11010045</li>
<li>Park, J. H., Cho, Y., Shin, D., Choi, S. (2022). Prediction of Mortality after Burn Surgery in Critically Ill Burn Patients Using Machine Learning Models. <em>Journal of Personalized Medicine</em>, <em>12</em>(8), 1293. https://doi.org/10.3390/jpm12081293</li>
<li>Feng, X., Feng, S. (2023). P1508: GUIDE PERSONALIZED ANTIBIOTIC STEWARDSHIP FOR PSEUDOMONAS AERUGINOSA BLOODSTREAM INFECTIONS IN AML PATIENTS: BASED ON A MACHINE LEARNING MODEL. <em>HemaSphere</em>, <em>7</em>(S3), e7972240. https://doi.org/10.1097/01.hs9.0000972912.79722.40</li>
<li>Unknown (1988). β-blockers may be more appropriate than diuretics for elderly hypertensive patients. <em>InPharma</em>, <em>655</em>(1), 3-4. https://doi.org/10.1007/bf03308910</li>
<li>Sadowski, B., Milewska, K., Ginter, J. (2022). Machine Learning Based Prediction of Gamma Passing Rate for VMAT Radiotherapy Plans. <em>Journal of Personalized Medicine</em>, <em>12</em>(12), 2071. https://doi.org/10.3390/jpm12122071</li>
<li>Bhowmik, R. T., Most, S. P. (2022). A Personalized Respiratory Disease Exacerbation Prediction Technique Based on a Novel Spatio-Temporal Machine Learning Architecture and Local Environmental Sensor Networks. <em>Electronics</em>, <em>11</em>(16), 2562. https://doi.org/10.3390/electronics11162562</li>
<li>Szolovits, P., Doyle, J., Long, W. J., Kohane, I. S., Pauker, S. G. (1994). Guardian Angel: Patient-Centered Health Information Systems. <em>DSpace@MIT (Massachusetts Institute of Technology)</em>.</li>
<li>Vita, A. D., Belmusto, A., Perna, F. D., Tremamunno, S., Matteis, G. D., Franceschi, F. (2024). The Impact of Climate Change and Extreme Weather Conditions on Cardiovascular Health and Acute Cardiovascular Diseases. <em>Journal of Clinical Medicine</em>, <em>13</em>(3), 759-759. https://doi.org/10.3390/jcm13030759</li>
<li>Doehner, W., Böhm, M., Boriani, G., Christersson, C., Coats, A. J., Hæusler, K. G. (2023). Interaction of Heart Failure and Stroke: A Clinical Consensus Statement of the ESC Council on Stroke, the Heart Failure Association (HFA) and the ESC Working Group on Thrombosis. <em>European Journal of Heart Failure</em>, <em>25</em>(12), 2107-2129. https://doi.org/10.1002/ejhf.3071</li>
<li>Nagamine, T., Gillette, B. M., Kahoun, J., Burghaus, R., Lippert, J., Saxena, M. (2022). Data-driven identification of heart failure disease states and progression pathways using electronic health records. <em>Scientific Reports</em>, <em>12</em>(1), 17871-17871. https://doi.org/10.1038/s41598-022-22398-4</li>
<li>Wang, T., Yan, Y., Xiang, S., Tan, J., Yang, C., Zhao, W. (2022). A comparative study of antihypertensive drugs prediction models for the elderly based on machine learning algorithms. <em>Frontiers in Cardiovascular Medicine</em>, <em>9</em>, 1056263-1056263. https://doi.org/10.3389/fcvm.2022.1056263</li>
<li>Lopes, M. A. C. Q., Oliveira, G. M. M. d., Ribeiro, A. L. P., Pinto, F. J., Rey, H. C. V., o, A. ́. A. ́. B. (2019). Guidelines os the Brazilian Society of Cardiology on Telemedicine in Cardiology - 2019. <em>Arquivos Brasileiros de Cardiologia</em>, <em>113</em>(5), 1006-1056. https://doi.org/10.5935/abc.20190205</li>
<li>Rohmann, N., Stürmer, P., Geisler, C., Schlicht, K., Knappe, C., Hartmann, K. (2023). Effects of lifestyle and associated diseases on serum CC16 suggest complex interactions among metabolism, heart and lungs. <em>Journal of Advanced Research</em>, <em>59</em>, 161-171. https://doi.org/10.1016/j.jare.2023.06.005</li>
<li>Herzog, L., Ber, R. I., Horowitz-Kugler, Z., Rabi, Y., Brufman, I., Paz, Y. E. (2023). Causal Deep Neural Network-Based Model for First-Line Hypertension Management. <em>Mayo Clinic Proceedings Digital Health</em>, <em>1</em>(4), 632-640. https://doi.org/10.1016/j.mcpdig.2023.10.001</li>
<li>Cristina, B., Helena, G. P., Ángel, O. Á. F., Fernando, S. R. (2017). Abstracts of the 50th Anniversary ESPN Meeting, Glasgow, September 2017. <em>Pediatric Nephrology</em>, <em>32</em>(9), 1643-1834. https://doi.org/10.1007/s00467-017-3753-x</li>
</ol>
</article>