Fatma Gunturkun
Data Science Manager, Med/Quantitative Sciences Unit
Current Role at Stanford
Data Science Manager
https://orcid.org/0000-0002-0948-0413All Publications
-
Evaluating Tumor Burden as a Predictive Biomarker for Epidermal Growth Factor Receptor Targeted Kinase Inhibitor Therapy in Advanced Non-Small Cell Lung Cancer.
JCO precision oncology
2026; 10 (4): e2500884
Abstract
As treatment options for advanced non-small cell lung cancer (NSCLC) evolve, biomarkers are needed to guide therapy selection while balancing efficacy and toxicity. Although tumor burden is a promising candidate, its prognostic role in guiding epidermal growth factor receptor (EGFR)-targeted kinase inhibitor (TKI) therapies remains understudied in real-world settings.We identified patients with de novo stage IV EGFR-mutant NSCLC treated with first-line EGFR-TKI at Stanford Health Care (2000-2021). Tumor burden metrics were manually annotated from 592 baseline radiology reports, encompassing size, number, and location (1,807 lesions). Multivariable Cox regression evaluated associations between tumor burden metric and overall survival (OS), adjusting for confounders, in the overall cohort and an osimertinib subgroup. A weighted composite tumor burden score was constructed using statistically significant metrics to stratify risk.Of 312 patients, bone metastasis (hazard ratio (HR)adjusted, 1.64 [95% CI, 1.23 to 2.19]) and the number of metastatic organs (HRadjusted, 1.21 [95% CI, 1.10 to 1.32]) were independently associated with worse OS and used to construct the composite score. Patients with low tumor burden (composite-score ≤ median 1.06) experienced better OS than those with high tumor burden, with a 3-year OS of 59.8% versus 41.5% (P = .001). Consistent findings were observed in the osimertinib subgroup, with a 3-year OS of 62.2% versus 44.6% (P = .03) for low versus high tumor burden.Tumor burden may serve as a prognostic biomarker in advanced NSCLC receiving EGFR-TKIs. These findings raise the hypothesis that durable survival in low-burden patients may be achievable with monotherapy, potentially sparing unnecessary toxicity from combination regimens. This warrants prospective validation comparing monotherapy versus combination strategies.
View details for DOI 10.1200/PO-25-00884
View details for PubMedID 41955549
-
Evaluation of Large Language Models as Emergency Department Revisit Predictors.
Pacific Symposium on Biocomputing. Pacific Symposium on Biocomputing
2026; 31: 130-143
Abstract
Large Language Models (LLMs) have shown promise in clinical reasoning and question answering, yet their effectiveness for real-world clinical prediction remains an open question. We present the first large-scale study evaluating LLMs for predicting 30-day emergency department (ED) revisits using 138,010 visits from the Adult Emergency Department at Stanford. We assessed two modeling paradigms: (1) direct prediction, where the LLM generates revisit risk assessments in natural language, and (2) embedding-based approaches that leverage LLM-derived vector representations (LLM2Vec) of patient data for downstream modeling. Retrieval augmentation improved direct prediction performance (e.g., Claude 3.7 F1 from 0.3755 (95% CI [0.3647, 0.3864]) to 0.4160 (95% CI [0.4024, 0.4294])), and embedding-based methods consistently outperformed direct approaches, with LLM2Vec achieving F1=0.4505 (95% CI [0.4345, 0.4666]). Despite having access to comprehensive structured and unstructured clinical data, all LLM approaches (F1=0.3022-0.4505) failed to exceed a traditional LightGBM model using only structured data (F1=0.4614 (95% CI [0.4496, 0.4789])). Through systematic analysis of the reasoning chains in 17,488 predictions, we suggest potential failure patterns: reasoning may systematically degrade performance through overweighting medical histories and similar visits, neglecting protective factors, and risk aversion. Our work establishes essential baseline performance while revealing fundamental limitations in current-generation LLMs for clinical prediction tasks.
View details for DOI 10.1142/9789819824755_0010
View details for PubMedID 41758138
-
ED-Explain: Personalized Video Instructions for Patients Discharged from the Emergency Department.
Pacific Symposium on Biocomputing. Pacific Symposium on Biocomputing
2026; 31: 280-293
Abstract
This paper presents ED-Explain, an integrated, AI-driven system that transforms emergency department (ED) discharge instructions and electronic health records into personalized video presentations featuring a virtual healthcare provider. By leveraging multimodal ED data, large language models, and video generation, we aimed to produce accessible discharge instructions tailored to patients' ED visits. Four board-certified Emergency Medicine physicians reviewed 39 pairs of original and ED-Explain-produced discharge instructions. AI video summaries received significantly higher (p<0.001) average ratings (1-5) of completeness (4.1 vs. 3.1), correctness (3.9 vs. 3.5) and patient accessibility (3.4 vs. 2.9). Physicians expressed reservations about 13.3% of ED-Explain's discharge instructions for patient viewing, and only 4.6% of ED-Explain's instructions were found inappropriate for use with patients. Physician feedback suggests that AI-enhanced video discharge summaries have potential to improve communication of discharge information to ED patients, though patient-centered evaluation is needed. This work contributes to the growing field of AIassisted healthcare communication and offers insights into the potential for AI to improve physician-patient communication and patient self-efficacy.
View details for DOI 10.1142/9789819824755_0020
View details for PubMedID 41758148
-
Digital Biometric Measures in Long COVID: A Secondary Analysis of the STOP-PASC Randomized Clinical Trial.
JAMA network open
2025; 8 (8): e2526901
Abstract
Digital biometrics can be used to monitor disease, but there is limited research on their applications for assessing postacute sequelae of SARS-CoV-2 (PASC) or long COVID.To better understand digital biometric patterns in long COVID using wearable device technology and whether there are any differences between the nirmatrelvir-ritonavir and placebo-ritonavir intervention arms.This secondary analysis is an exploratory substudy of a placebo-controlled randomized clinical trial (Selective Trial of Paxlovid for PASC [STOP-PASC]) that was conducted from November 2022 to September 2023 at Stanford University. Trial participants were randomized, and a subset enrolled into the prespecified substudy.Participants were randomized 2:1 to receive oral nirmatrelvir (300 mg) and ritonavir (100 mg) or placebo-ritonavir twice daily for 15 days. Substudy participants were provided with a smartwatch and asked to wear it for 24 hours a day, 7 days a week for the 15-week study.Mean changes in digital biometric measures in physical activity, heart rate, and oxygen saturation tracked by a smartwatch. Biometric measures were summarized by treatment arm for each of 5 different time frames: baseline, treatment period, and 3 subsequent time intervals. Summary measure trajectories were clustered and demographics and clinical characteristics were compared among clusters using absolute standardized differences expressed in units of SDs.Of the 94 participants enrolled in the substudy, 50 (37 in nirmatrelvir-ritonavir and 13 in placebo-ritonavir) met the analysis eligibility criteria based on wear time and data completeness. These participants had a mean (SD) age of 42.7 (13.2) years and included 29 females (58.0%). Using mixed models for repeated measures, no significant differences were detected between the intervention arms in the change in biometric measures over time, consistent with the patient-reported outcomes in the STOP-PASC trial. In the overall substudy cohort, latent class mixed models and cluster analysis identified distinct longitudinal trajectories of long COVID over the 15-week study that tracked with different symptoms. Participants with lower daytime physical activity reported more severe fatigue (9 of 9 [100%] vs 21 of 23 [91.3%]; absolute standardized difference [ASD], 0.30), shortness of breath (9 of 9 [100%] vs 7 of 23 [30.4%]; ASD, 1.31), and cardiovascular symptoms (8 of 9 [88.9%] vs 12 of 23 [52.2%]; ASD, 0.70). Those with higher nighttime physical activity reported more gastrointestinal symptoms (2 of 3 [66.7%] vs 11 of 35 [31.4%]; ASD, 0.50). Additionally, participants with higher median daytime heart rates reported less fatigue (7 of 9 [77.8%] vs 39 of 40 [97.5%]; ASD, 0.63) and shortness of breath (3 of 9 [33.3%] vs 23 of 40 [57.5%]; ASD, 0.50) compared with those with lower heart rates.This secondary analysis identified distinct longitudinal trajectories of physiological and behavioral digital biometric measures captured from wearable devices that reflect the heterogeneity and track with different symptoms of long COVID. Digital biometric measures from wearable devices have promising utility for long COVID research.ClinicalTrials.gov Identifier: NCT05576662.
View details for DOI 10.1001/jamanetworkopen.2025.26901
View details for PubMedID 40810942
View details for PubMedCentralID PMC12355292
-
The impact of pregnancy on mortality and lung function in cystic fibrosis patients.
Journal of cystic fibrosis : official journal of the European Cystic Fibrosis Society
2025
Abstract
As the lifespan of people with cystic fibrosis (pwCF) improves, more individuals are pursuing pregnancy. Historically, pregnancy was not recommended in this population; however, more recent evidence has revealed inconsistent survival and lung function outcomes. Our aim was to assess the differences in survival and lung function between pregnant and never-pregnant pwCF and to provide updated recommendations for contemporary clinical practice.In this retrospective matched parallel cohort study, data was collected from the American Cystic Fibrosis Foundation Patient Registry (CFFPR) from 1999 to 2019. 1743 adult pwCF with a reported pregnancy were matched with 1743 never-pregnant patients. Regression models were developed to estimate associations between patient characteristics, pregnancy, and outcomes. The primary endpoint was the probability of survival comparing pregnant and never-pregnant pwCF, while the secondary endpoint was lung function over time.The study cohort (n = 3486) had a mean age of 24.96 years. There was no significant difference in survival probabilities between pregnant and never-pregnant pwCF (56.2 %, CI95 %: 51.3 %-61.5 % vs. 55.8 %, CI95 %: 52.1 %-59.7 %, p = 0.5). The multivariable time-dependent Cox regression analysis resulted in a significantly lower mortality hazard rate for pregnant cohorts (HR:0.78, p < 0.01). There was no significant association between pregnancy and lung function over time (0.99, p = 0.21).Pregnancy was associated with a reduced hazard of death compared to never-pregnant pwCF and did not demonstrate a significant impact on lung function. Therefore, pregnancy should not be generally discouraged in pwCF and clinicians should evaluate pregnancy risks and benefits on an individualized basis.
View details for DOI 10.1016/j.jcf.2025.02.004
View details for PubMedID 39984374
-
Applying Large Language Models to Assess Quality of Care: Monitoring ADHD Medication Side Effects.
Pediatrics
2025; 155 (1)
Abstract
To assess the accuracy of a large language model (LLM) in measuring clinician adherence to practice guidelines for monitoring side effects after prescribing medications for children with attention-deficit/hyperactivity disorder (ADHD).Retrospective population-based cohort study of electronic health records. Cohort included children aged 6 to 11 years with ADHD diagnosis and 2 or more ADHD medication encounters (stimulants or nonstimulants prescribed) between 2015 and 2022 in a community-based primary health care network (n = 1201). To identify documentation of side effects inquiry, we trained, tested, and deployed an open-source LLM (LLaMA) on all clinical notes from ADHD-related encounters (ADHD diagnosis or ADHD medication prescription), including in-clinic/telehealth and telephone encounters (n = 15 628 notes). Model performance was assessed using holdout and deployment test sets, compared with manual medical record review.The LLaMA model accurately classified notes that contained side effects inquiry (sensitivity = 87.2, specificity = 86.3, area under curve = 0.93 on holdout test set). Analyses revealed no model bias in relation to patient sex or insurance. Mean age (SD) at first prescription was 8.8 (1.6) years; characteristics were mostly similar across patients with and without documented side effects inquiry. Rates of documented side effects inquiry were lower for telephone encounters than for in-clinic/telehealth encounters (51.9% vs 73.0%, P < .001). Side effects inquiry was documented in 61.4% of encounters after stimulant prescriptions and 48.5% of encounters after nonstimulant prescriptions (P = .041).Deploying an LLM on a variable set of clinical notes, including telephone notes, offered scalable measurement of quality of care and uncovered opportunities to improve psychopharmacological medication management in primary care.
View details for DOI 10.1542/peds.2024-067223
View details for PubMedID 39701141
-
Mental health care-seeking and barriers: a cross-sectional study of an urban Latinx community.
BMC public health
2024; 24 (1): 3091
Abstract
The Latinx community faces an increasing amount of mental health challenges and disparities in care. While the contributing factors are complex, there are likely potential barriers related to connecting with mental health support and accessing care that can be addressed.To investigate barriers in connecting to mental health care, we conducted a cross-sectional survey of mental health service use and barriers in an urban community with a primarily Hispanic/Latinx ethnicity using a modified random walk approach for door-to-door data collection with a two-cluster sampling frame. Survey included questions on socio-demographic, mental health status, desire and attempt to seek care, and the Barriers to Access to Care Evaluation. Shapley additive explanation (SHAP) identified impactful barriers and demographic characteristics. Our primary outcome was the number of respondents who saw a professional in the past 12 months and the key determinants that enabled their successful connection. Secondary outcomes were people with poor mental health who had wanted or tried to seek any source of mental health support.Of the 1004 respondents enrolled, 70.5% were foreign born; 63.4% were women. In the past 12 months, 23.8% of respondents wanted to connect with mental health care; 15.5% tried to connect, and only 11.7% successfully connected to mental health services. The two most cited barriers had the highest SHAP values: concerns about treatments available (65%) and financial costs (62.7%). Additional barriers with high SHAP values: being seen as weak and having no one to help them find care. Of demographic characteristics, age had the highest SHAP values.In a community with a high density of Latinx immigrants, just under half of respondents wanting mental health care successfully connected. Perceived informational, financial, and stigma-related barriers impacted the likelihood to connect with mental health care. These factors should be considered when designing programs and interventions to improve mental health care access and services in the Latinx community.
View details for DOI 10.1186/s12889-024-20533-6
View details for PubMedID 39516848
View details for PubMedCentralID 10413745
-
Advancing presurgical non-invasive molecular subgroup prediction in medulloblastoma using artificial intelligence and MRI signatures.
Cancer cell
2024
Abstract
Global investigation of medulloblastoma has been hindered by the widespread inaccessibility of molecular subgroup testing and paucity of data. To bridge this gap, we established an international molecularly characterized database encompassing 934 medulloblastoma patients from thirteen centers across China and the United States. We demonstrate how image-based machine learning strategies have the potential to create an alternative pathway for non-invasive, presurgical, and low-cost molecular subgroup prediction in the clinical management of medulloblastoma. Our robust validation strategies-including cross-validation, external validation, and consecutive validation-demonstrate the model's efficacy as a generalizable molecular diagnosis classifier. The detailed analysis of MRI characteristics replenishes the understanding of medulloblastoma through a nuanced radiographic lens. Additionally, comparisons between East Asia and North America subsets highlight critical management implications. We made this comprehensive dataset, which includes MRI signatures, clinicopathological features, treatment variables, and survival data, publicly available to advance global medulloblastoma research.
View details for DOI 10.1016/j.ccell.2024.06.002
View details for PubMedID 38942025
-
Machine learning evaluation of inequities and disparities associated with nurse sensitive indicator safety events.
Journal of nursing scholarship : an official publication of Sigma Theta Tau International Honor Society of Nursing
2024
Abstract
PURPOSE: To use machine learning to examine health equity and clinical outcomes in patients who experienced a nurse sensitive indicator (NSI) event, defined as a fall, a hospital-acquired pressure injury (HAPI) or a hospital-acquired infection (HAI).DESIGN: This was a retrospective observational study from a single academic hospital over six calendar years (2016-2021). Machine learning was used to examine patients with an NSI compared to those without.METHODS: Inclusion criteria: all adult inpatient admissions (2016-2021). Three approaches were used to analyze the NSI group compared to the No-NSI group. In the univariate analysis, descriptive statistics, and absolute standardized differences (ASDs) were employed to compare the demographics and clinical variables of patients who experienced a NSI and those who did not experience any NSIs. For the multivariate analysis, a light grading boosting machine (LightGBM) model was utilized to comprehensively examine the relationships associated with the development of an NSI. Lastly, a simulation study was conducted to quantify the strength of associations obtained from the machine learning model.RESULTS: From 163,507 admissions, 4643 (2.8%) were associated with at least one NSI. The mean, standard deviation (SD) age was 59.5 (18.2) years, males comprised 82,397 (50.4%). Non-Hispanic White 84,760 (51.8%), non-Hispanic Black 8703 (5.3%), non-Hispanic Asian 23,368 (14.3%), non-Hispanic Other 14,284 (8.7%), and Hispanic 30,271 (18.5%). Race and ethnicity alone were not associated with occurrence of an NSI. The NSI group had a statistically significant longer length of stay (LOS), longer intensive care unit (ICU) LOS, and was more likely to have an emergency admission compared to the group without an NSI. The simulation study results demonstrated that likelihood of NSI was higher in patients admitted under the major diagnostic categories (MDC) associated with circulatory, digestive, kidney/urinary tract, nervous, and infectious and parasitic disease diagnoses.CONCLUSION: In this study, race/ethnicity was not associated with the risk of an NSI event. The risk of an NSI event was associated with emergency admission, longer LOS, longer ICU-LOS and certain MDCs (circulatory, digestive, kidney/urinary, nervous, infectious, and parasitic diagnoses).CLINICAL RELEVANCE: Machine learning methodologies provide a new mechanism to investigate NSI events through the lens of health equity/disparity. Understanding which patients are at higher risk for adverse outcomes can help hospitals improve nursing care and prevent NSI injury and harm.
View details for DOI 10.1111/jnu.12983
View details for PubMedID 38773783
-
AI-based preeclampsia detection and prediction with electrocardiogram data.
Frontiers in cardiovascular medicine
2024; 11: 1360238
Abstract
More than 76,000 women die yearly from preeclampsia and hypertensive disorders of pregnancy. Early diagnosis and management of preeclampsia can improve outcomes for both mother and baby. In this study, we developed artificial intelligence models to detect and predict preeclampsia from electrocardiograms (ECGs) in point-of-care settings.Ten-second 12-lead ECG data was obtained from two large health care settings: University of Tennessee Health Science Center (UTHSC) and Atrium Health Wake Forest Baptist (AHWFB). UTHSC data was split into 80% training and 20% holdout data. The model used a modified ResNet convolutional neural network, taking one-dimensional raw ECG signals comprising 12 channels as an input, to predict risk of preeclampsia. Sub-analyses were performed to assess the predictive accuracy for preeclampsia prediction within 30, 60, or 90 days before diagnosis.The UTHSC cohort included 904 ECGs from 759 females (78.8% African American) with a mean ± sd age of 27.3 ± 5.0 years. The AHWFB cohort included 817 ECGs from 141 females (45.4 African American) with a mean ± sd age of 27.4 ± 5.9 years. The cross-validated ECG-AI model yielded an AUC (95% CI) of 0.85 (0.77-0.93) on UTHSC holdout data, and an AUC (95% CI) of 0.81 (0.77-0.84) on AHWFB data. The sub-analysis of different time windows before preeclampsia prediction resulted in AUCs (95% CI) of 0.92 (0.84-1.00), 0.89 (0.81-0.98) and 0.90 (0.81-0.98) when tested on ECGs 30 days, 60 days and 90 days, respectively, before diagnosis. When assessed on early onset preeclampsia (preeclampsia diagnosed at <34 weeks of pregnancy), the model's AUC (95% CI) was 0.98 (0.89-1.00).We conclude that preeclampsia can be identified with high accuracy via application of AI models to ECG data.
View details for DOI 10.3389/fcvm.2024.1360238
View details for PubMedID 38500752
View details for PubMedCentralID PMC10945012
-
Externally validated deep learning model to identify prodromal Parkinson's disease from electrocardiogram.
Scientific reports
2023; 13 (1): 12290
Abstract
Little is known about electrocardiogram (ECG) markers of Parkinson's disease (PD) during the prodromal stage. The aim of the study was to build a generalizable ECG-based fully automatic artificial intelligence (AI) model to predict PD risk during the prodromal stage, up to 5 years before disease diagnosis. This case-control study included samples from Loyola University Chicago (LUC) and University of Tennessee-Methodist Le Bonheur Healthcare (MLH). Cases and controls were matched according to specific characteristics (date, age, sex and race). Clinical data were available from May, 2014 onward at LUC and from January, 2015 onward at MLH, while the ECG data were available as early as 1990 in both institutes. PD was denoted by at least two primary diagnostic codes (ICD9 332.0; ICD10 G20) at least 30 days apart. PD incidence date was defined as the earliest of first PD diagnostic code or PD-related medication prescription. ECGs obtained at least 6 months before PD incidence date were modeled to predict a subsequent diagnosis of PD within three time windows: 6 months-1 year, 6 months-3 years, and 6 months-5 years. We applied a novel deep neural network using standard 10-s 12-lead ECGs to predict PD risk at the prodromal phase. This model was compared to multiple feature engineering-based models. Subgroup analyses for sex, race and age were also performed. Our primary prediction model was a one-dimensional convolutional neural network (1D-CNN) that was built using 131 cases and 1058 controls from MLH, and externally validated on 29 cases and 165 controls from LUC. The model was trained on 90% of the MLH data, internally validated on the remaining 10% and externally validated on LUC data. The best performing model resulted in an external validation AUC of 0.67 when predicting future PD at any time between 6 months and 5 years after the ECG. Accuracy increased when restricted to ECGs obtained within 6 months to 3 years before PD diagnosis (AUC 0.69) and was highest when predicting future PD within 6 months to 1 year (AUC 0.74). The 1D-CNN model based on raw ECG data outperformed multiple models built using more standard ECG feature engineering approaches. These results demonstrate that a predictive model developed in one cohort using only raw 10-s ECGs can effectively classify individuals with prodromal PD in an independent cohort, particularly closer to disease diagnosis. Standard ECGs may help identify individuals with prodromal PD for cost-effective population-level early detection and inclusion in disease-modifying therapeutic trials.
View details for DOI 10.1038/s41598-023-38782-7
View details for PubMedID 37516770
View details for PubMedCentralID PMC10387090
-
Development of a Deep Learning Model for Retinal Hemorrhage Detection on Head Computed Tomography in Young Children.
JAMA network open
2023; 6 (6): e2319420
Abstract
Abusive head trauma (AHT) in children is often missed in medical encounters, and retinal hemorrhage (RH) is considered strong evidence for AHT. Although head computed tomography (CT) is obtained routinely, all but exceptionally large RHs are undetectable on CT images in children.To examine whether deep learning-based image analysis can detect RH on pediatric head CT.This diagnostic study included 301 patients diagnosed with AHT who underwent head CT and dilated fundoscopic examinations at a quaternary care children's hospital. The study assessed a deep learning model using axial slices from 218 segmented globes with RH and 384 globes without RH between May 1, 2007, and March 31, 2021. Two additional light gradient boosting machine (GBM) models were assessed: one that used demographic characteristics and common brain findings in AHT and another that combined the deep learning model's risk prediction plus the same demographic characteristics and brain findings.Sensitivity (recall), specificity, precision, accuracy, F1 score, and area under the curve (AUC) for each model predicting the presence or absence of RH in globes were assessed. Globe regions that influenced the deep learning model predictions were visualized in saliency maps. The contributions of demographic and standard CT features were assessed by Shapley additive explanation.The final study population included 301 patients (187 [62.1%] male; median [range] age, 4.6 [0.1-35.8] months). A total of 120 patients (39.9%) had RH on fundoscopic examinations. The deep learning model performed as follows: sensitivity, 79.6%; specificity, 79.2%; positive predictive value (precision), 68.6%; negative predictive value, 87.1%; accuracy, 79.3%; F1 score, 73.7%; and AUC, 0.83 (95% CI, 0.75-0.91). The AUCs were 0.80 (95% CI, 0.69-0.91) for the general light GBM model and 0.86 (95% CI, 0.79-0.93) for the combined light GBM model. Sensitivities of all models were similar, whereas the specificities of the deep learning and combined light GBM models were higher than those of the light GBM model.The findings of this diagnostic study indicate that a deep learning-based image analysis of globes on pediatric head CTs can predict the presence of RH. After prospective external validation, a deep learning model incorporated into CT image analysis software could calibrate clinical suspicion for AHT and provide decision support for which patients urgently need fundoscopic examinations.
View details for DOI 10.1001/jamanetworkopen.2023.19420
View details for PubMedID 37347482
View details for PubMedCentralID PMC10288337
-
Artificial intelligence opportunities in cardio-oncology: Overview with spotlight on electrocardiography
AMERICAN HEART JOURNAL PLUS: CARDIOLOGY RESEARCH AND PRACTICE
2022; 15
View details for DOI 10.1016/j.ahjo.2022.100129
View details for Web of Science ID 001133755300005
-
Artificial Intelligence-Assisted Prediction of Late-Onset Cardiomyopathy Among Childhood Cancer Survivors.
JCO clinical cancer informatics
2021; 5: 459-468
Abstract
Early identification of childhood cancer survivors at high risk for treatment-related cardiomyopathy may improve outcomes by enabling intervention before development of heart failure. We implemented artificial intelligence (AI) methods using the Children's Oncology Group guideline-recommended baseline ECG to predict cardiomyopathy.Seven AI and signal processing methods were applied to 10-second 12-lead ECGs obtained on 1,217 adult survivors of childhood cancer prospectively followed in the St Jude Lifetime Cohort (SJLIFE) study. Clinical and echocardiographic assessment of cardiac function was performed at initial and follow-up SJLIFE visits. Cardiomyopathy was defined as an ejection fraction < 50% or an absolute drop from baseline ≥ 10%. Genetic algorithm was used for feature selection, and extreme gradient boosting was applied to predict cardiomyopathy during the follow-up period. Model performance was evaluated by five-fold stratified cross-validation.The median age at baseline SJLIFE evaluation was 31.7 years (range 18.4-66.4), and the time between baseline and follow-up evaluations was 5.2 years (0.5-9.5). Two thirds (67.1%) of patients were exposed to chest radiation, and 76.6% to anthracycline chemotherapy. One hundred seventeen (9.6%) patients developed cardiomyopathy during follow-up. In the model based solely on ECG features, the cross-validation area under the curve (AUC) was 0.87 (95% CI, 0.83 to 0.90), whereas the model based on clinical features had an AUC of 0.69 (95% CI, 0.64 to 0.74). In the model based on ECG and clinical features, the cross-validation AUC was 0.89 (95% CI, 0.86 to 0.91), with a sensitivity of 78% and a specificity of 81%.AI using ECG data may assist in the identification of childhood cancer survivors at increased risk for developing future cardiomyopathy.
View details for DOI 10.1200/CCI.20.00176
View details for PubMedID 33909450
View details for PubMedCentralID PMC8462657
-
Deep learning for improved prediction of late-onset cardiomyopathy among childhood cancer survivors: A report from the St. Jude Lifetime Cohort (SJLIFE).
LIPPINCOTT WILLIAMS & WILKINS. 2020
View details for Web of Science ID 000560368300263