Enhancing Cause of Death Prediction: Development and Validation of ML Models Using Multimodal Data Across Multiple Healthcare Sites

Enhancing Cause of Death Prediction: Development and Validation of ML Models Using Multimodal Data Across Multiple Healthcare Sites | medRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-P4HH5NV'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search Enhancing Cause of Death Prediction: Development and Validation of ML Models Using Multimodal Data Across Multiple Healthcare Sites View ORCID Profile Mohammed Al-Garadi , Rishi J Desai , Kerry Ngan , Michele LeNoue-Newton , Ruth M. Reeves , Daniel Park , Jose J. Hernández-Muñoz , Shirley V. Wang , Judith C. Maro , Candace C. Fuller , Joshua Lin Kueiyu , Aida Kuzucan , Kevin Coughlin , Haritha Pillai , Melissa McPheeters , Jill Whitaker , Jessica A. Deere , Michael F. McLemore , Dax M. Westerman , Tony Morrow , Margaret A. Adgent , Michael E. Matheny doi: https://doi.org/10.1101/2025.06.24.25330213 Mohammed Al-Garadi 1 Department of Biomedical Informatics, Vanderbilt University Medical Center , Nashville, TN, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Mohammed Al-Garadi For correspondence: mohammed.a.al-garadi{at}vumc.org Rishi J Desai 3 Division of Pharmacoepidemiology and Pharmacoeconomics, Department of Medicine, Brigham and Women’s Hospital and Harvard Medical School , Boston, Massachusetts, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site Kerry Ngan 3 Division of Pharmacoepidemiology and Pharmacoeconomics, Department of Medicine, Brigham and Women’s Hospital and Harvard Medical School , Boston, Massachusetts, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site Michele LeNoue-Newton 1 Department of Biomedical Informatics, Vanderbilt University Medical Center , Nashville, TN, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site Ruth M. Reeves 1 Department of Biomedical Informatics, Vanderbilt University Medical Center , Nashville, TN, USA 2 Geriatrics Research Education and Clinical Care Service & VINCI, Tennessee Valley Healthcare System VA , Nashville, TN, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site Daniel Park 1 Department of Biomedical Informatics, Vanderbilt University Medical Center , Nashville, TN, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site Jose J. Hernández-Muñoz 6 Office of Surveillance and Epidemiology, Center for Drug Evaluation and Research, US Food and Drug Administration , Silver Spring, MD Find this author on Google Scholar Find this author on PubMed Search for this author on this site Shirley V. Wang 3 Division of Pharmacoepidemiology and Pharmacoeconomics, Department of Medicine, Brigham and Women’s Hospital and Harvard Medical School , Boston, Massachusetts, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site Judith C. Maro 4 Harvard Pilgrim Health Care Institute and Department of Population Medicine, Harvard Medical School , Boston, MA, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site Candace C. Fuller 4 Harvard Pilgrim Health Care Institute and Department of Population Medicine, Harvard Medical School , Boston, MA, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site Joshua Lin Kueiyu 3 Division of Pharmacoepidemiology and Pharmacoeconomics, Department of Medicine, Brigham and Women’s Hospital and Harvard Medical School , Boston, Massachusetts, USA 5 Department of Medicine, Massachusetts General Hospital and Harvard Medical 0School , Boston, MA, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site Aida Kuzucan 6 Office of Surveillance and Epidemiology, Center for Drug Evaluation and Research, US Food and Drug Administration , Silver Spring, MD Find this author on Google Scholar Find this author on PubMed Search for this author on this site Kevin Coughlin 4 Harvard Pilgrim Health Care Institute and Department of Population Medicine, Harvard Medical School , Boston, MA, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site Haritha Pillai 3 Division of Pharmacoepidemiology and Pharmacoeconomics, Department of Medicine, Brigham and Women’s Hospital and Harvard Medical School , Boston, Massachusetts, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site Melissa McPheeters 7 RTI International Find this author on Google Scholar Find this author on PubMed Search for this author on this site Jill Whitaker 1 Department of Biomedical Informatics, Vanderbilt University Medical Center , Nashville, TN, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site Jessica A. Deere 1 Department of Biomedical Informatics, Vanderbilt University Medical Center , Nashville, TN, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site Michael F. McLemore 1 Department of Biomedical Informatics, Vanderbilt University Medical Center , Nashville, TN, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site Dax M. Westerman 1 Department of Biomedical Informatics, Vanderbilt University Medical Center , Nashville, TN, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site Tony Morrow 8 Department of Health Policy, Vanderbilt University Medical Center , Nashville, TN Find this author on Google Scholar Find this author on PubMed Search for this author on this site Margaret A. Adgent 8 Department of Health Policy, Vanderbilt University Medical Center , Nashville, TN Find this author on Google Scholar Find this author on PubMed Search for this author on this site Michael E. Matheny 1 Department of Biomedical Informatics, Vanderbilt University Medical Center , Nashville, TN, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site Abstract Full Text Info/History Metrics Data/Code Preview PDF Abstract Importance Timely and accurate determination of causes of death (CoD) is essential for public health surveillance, epidemiological research, and healthcare policy development. However, obtaining up-to-date and detailed CoD information is challenging due to delays in official death records and inconsistencies in data reporting across institutions. Objective To develop and validate machine learning (ML) models capable of predicting probable CoD by integrating comprehensive features from structured electronic health record (EHR) data, unstructured clinical notes, and publicly available data. Design, Setting, and Participants This multi-institutional retrospective cohort study was conducted at Vanderbilt University Medical Center (VUMC) and Massachusetts General Brigham (MGB). Deceased patients were included if they had at least one inpatient or outpatient encounter between October 1, 2015, and January 1, 2021, with corresponding death records from state health departments and the National Death Index. The study was comprised of 13,708 deceased patients from VUMC and 34,839 from MGB. Exposures Integration of structured EHR data, unstructured clinical notes processed using advanced language models, and publicly available data into machine learning models to predict CoD. Main Outcomes and Measures The primary outcome was the underlying CoD, classified into one of the top 15 National Center for Health Statistics (NCHS) rankable CoD categories, with all other causes grouped into an “Other” category. Model performance was evaluated using weighted area under the receiver operating characteristic curve (AUC) and weighted F-measure. Results The XGBoost model using structured EHR data alone achieved weighted AUCs of 0.86 (95% CI, 0.84–0.88) at VUMC and 0.80 (95% CI, 0.79-0.80) at MGB. Adding unstructured notes improved performance, with weighted AUCs of 0.90 (95% CI, 0.88–0.93) at VUMC and 0.92(95% CI, 0.91–0.92) at MGB. Adding publicly available data did not further improve performance. Cross-institutional validation revealed significant performance degradation. Conclusions and Relevance ML models integrating EHR structured and unstructured data to predict underlying CoD at the time of the most recent encounter among deceased patients achieved excellent performance within individual institutions. The inclusion of publicly available data did not improve performance, and all versions had poor portability between institutions. Healthcare institutions may benefit from adopting robust processes for locally tailored models, and future research should focus on enhancing model generalizability while addressing unique institutional data environments. Introduction Timely and accurate identification of death and determination of causes of death (CoD) are vital for public health surveillance, epidemiological research, and healthcare policy development. Mortality is a critical outcome in medical product safety and effectiveness assessments, with all-cause mortality being one of the most studied outcomes in healthcare. 1 – 4 Such mortality data inform strategies for disease prevention, resource allocation, and the evaluation of healthcare interventions. Obtaining up-to-date and detailed CoD information presents significant challenges due to delays in official death records and inconsistencies in data reporting across institutions. State-level mortality data can be delayed by 1 to 1.5 years, while national mortality data may lag by up to 2 years because of the time required for data collection, processing, and validation. 5 , 6 This delay restricts its utility for analyses requiring more recent data ascertainment. 5 , 6 Electronic health records (EHRs) offer a promising solution by providing a rich repository of patient information, including structured data—such as diagnoses, procedures, medications, laboratory results, and vital signs—and unstructured data like clinical notes. Leveraging EHR data for CoD prediction holds significant potential but presents challenges due to the heterogeneity of data sources and the complexity of integrating diverse data types. 7 Previous studies to predict CoD using EHR data have often been limited by single-site data, smaller populations, or a focus on specific CoD, which may not generalize to broader, more diverse populations. 8 – 13 For instance, Kim et al. 9 used a stacking ensemble method to predict CoD from patient checkup data, but their approach was limited by the exclusion of unstructured data and challenges in expanding to other institutional datasets. 9 Jeblee et al. addressed CoD prediction using verbal autopsy (VA) narratives, applying natural language processing (NLP) techniques to unstructured narrative data. Their method utilized word frequency counts from free-text narratives to improve CoD classification accuracy, achieving 77% sensitivity for adult deaths across 15 categories. While their narrative-based approach outperformed structured data classifiers at the individual level, it was still limited to population-level estimations in low-resource settings. 8 Additionally, the model’s reliance on the VA structure hindered its applicability to broader clinical datasets. Our study represents a significant advancement by integrating both structured EHR data and unstructured clinical notes using advanced language models, resulting in a more comprehensive and scalable prediction system than those in previous studies. The inclusion of multiple multimodal data sources—including EHRs, clinical notes, and publicly available records—establishes a level of data integration not previously achieved, thereby extending the scope beyond prior research. Moreover, our study’s multi-institutional design encompasses a large and diverse patient population, thereby strengthening the analysis and conclusions drawn from extensive data. 8 – 13 Integrating unstructured clinical narratives using advanced language model techniques can capture nuanced clinical information that is not readily available in structured data. 14 – 17 Additionally, publicly available data sources, such as obituaries and memorial websites, may provide supplementary information that enhances mortality surveillance. 18 – 20 Incorporating these varied data sources requires sophisticated feature engineering and validation against standardized reference-standard labels to ensure accuracy and reliability. Recent advances in deep learning and NLP, especially transformer-based models like BERT and GPT, have improved the integration of unstructured data in healthcare. 21 These models effectively process large volumes of clinical text, capturing complex linguistic patterns and medical terminology. By leveraging these advanced NLP and large language model (LLM) techniques, healthcare institutions can more readily incorporate unstructured data into predictive models, enhancing accuracy without extensive manual feature engineering. 22 , 23 This scalability accelerates the development of predictive tools, making them more accessible for widespread clinical application and ultimately improving patient outcomes through informed decision-making. 23 , 24 In this multi-institutional retrospective cohort study, we aimed to develop and validate machine learning models capable of predicting CoD by integrating comprehensive features from structured EHR data, unstructured clinical notes, and publicly available data. We hypothesized that integrating multimodal data sources with advanced NLP techniques would improve CoD prediction accuracy compared to models using structured data alone. Conducted at Vanderbilt University Medical Center (VUMC) and Massachusetts General Brigham (MGB), our study included a large patient population with a wide range of CoD. By utilizing standardized reference-standard death records from the National Death Index (NDI) and state health departments, we ensured reliable labels for model training and evaluation. Methods Study Design This multi-institutional, retrospective cohort study aimed to develop and validate machine learning models for predicting CoD by integrating structured EHR data, unstructured EHR clinical notes, and publicly available data ( Figure 1 ). Conducted at VUMC and MGB, the pipeline, as illustrated in Figure 1 , incorporated feature engineering from structured EHR data and document-level embeddings from clinical notes using the Clinical Longformer model. Public data sources were categorized into top CoD groups and one-hot encoded with a large language model. These inputs were combined into machine learning models, trained using gold-standard death records from state health departments and the NDI, and cross-evaluated across both institutions for portability capacity and generalizability. Download figure Open in new tab Figure 1: Pipeline for CoD Prediction Using Multimodal Data Sources and Machine Learning Models Study Population We included retrospective cohorts from both MBG and VUMC in this study. For MGB, patients whose deaths were recorded between January 1, 2015, and Dec 31, 2020, in one of 3 state vital statistic database (Massachusetts, Connecticut, and Vermont) were eligible if they had ≥1 visit recorded in MGB EHRs within 180 days prior to death. Inclusion also required at least one clinical note within a six-month look-back period prior to the last healthcare system encounter. For VUMC, patients were included if they had at least one inpatient or outpatient encounter within the same date range, matched to the NDI with a recorded date of death between January 1, 2019, and December 31, 2021. They were also required to have at least one inpatient or outpatient encounter documented in the EHR within the six-month look-back period prior to their last recorded encounter. CoD Outcome Definitions The outcomes of interest were the causes of death. These data are reported as part of the death certificate submissions in each state following the death of the patient. To align with how vital statistics data are commonly reported and used in observational research, we classified CoD into the 15 most frequent National Center for Health Statistics (NCHS) rankable categories, with all other causes grouped into an “Other” category. 25 Reference Standard Data Collection The reference standard for CoD was the official death record from the NDI for VUMC and the state health departments of Massachusetts, Connecticut, and Vermont for MGB. These records provided definitive and reliable sources for determining the primary CoD. 26 , 27 The term “reference -standard death records” refers specifically to this authoritative mortality data from the NDI and state health departments. Data Sources & Feature Engineering We collected data elements from the EHRs of both VUMC and MGB. Table 1 summarizes the model’s integrated data layers: structured EHR data (ICD codes, labs, meds, vitals), unstructured EHR notes (language model embeddings), and public data (top causes of death from the National Center for Health Statistics (NCHS) 25 groupers). View this table: View inline View popup Table 1 provides a summary of the data layers and features integrated into the model. First, structured data elements were extracted that included diagnoses, procedures, medications, laboratory orders and results, vital signs, and demographic information. We transformed these sparse data into clinically meaningful groupings to reduce sparsity. Accordingly, we mapped ICD-10-CM diagnosis codes and ICD-10-PCS procedure codes to the Clinical Classification Software categories. 28 Next, we grouped laboratory tests using the Regenstreif Logical Observation Identifiers Names and Codes [LOINC] test grouper mappings. 29 We then transformed the medications into a RxNorm ingredient level binary yes/no exposure at the time of the last encounter and aggregated these data to the Anatomical Therapeutic Chemical [ATC] classification level 4. 30 We also included the following vital signs: oxygen saturation, temperature, systolic and diastolic blood pressure, respiratory rate, and heart rate), and demographic data (Race, Gender, Ethnicity). This resulted in a total of 2,048 structured features. Second, unstructured data comprising clinical notes were collected from the six months preceding the last patient encounter. We derived 768 unstructured features from embeddings of clinical notes using the Clinical Longformer small language model, 31 which can process up to 4,096 tokens per document. Third, we collected publicly available data from obituaries, crowdfunding platforms like GoFundMe, and memorial websites: Everloved and TributeArchives. A summary of the data characteristics of these data sources is noted in our previous paper. 32 We derived 16 binary public data features representing the top 15 CoD categories defined by the NCHS 25 and an “Other” category. CoD from public data was extracted and categorized using the LLaMA 3 LLM, leveraging our developed model to ensure alignment with standardized definitions. 32 Data Source Linkage All data sources were linked using unique patient identifiers, resulting in an integrated dataset that included three types of features. The integration of public data with each institution’s EHR employed a method described in our previous paper. 32 Public data, such as obituaries and memorial websites, were first downloaded and cleaned offline to ensure the data was processed in a controlled environment. The linkage between public data and EHR data was conducted on a local, secure server without exposure of the data to the internet, following strict data privacy protocols. We matched decedent information extracted from social media sources to patient records in the EHR using probabilistic linkage algorithms (SPLINK) based on first name, last name, date of birth, and state of residence. The linkage process demonstrated high accuracy, with exact matches on these variables indicating the reliable performance of the matching methodology. 33 By linking and managing all data within each healthcare institution, we adhered to both institutional standards for patient privacy. Feature Selection To enhance model interpretability and prevent overfitting, we applied feature selection techniques. Stability selection 34 utilized base estimators like Lasso 35 or Random Forest 36 to identify features consistently associated with the outcome across subsamples. Random Forest feature selection involves extracting feature importance scores from trained Random Forest models. Additionally, we employed SelectFromModel using XGBoost’s feature importance metrics to retain the top features. 37 Model Development At each site, ML models were developed to predict the top 15 underlying CoD categories and evaluated stepwise for three levels of feature inclusion (EHR structured data, EHR structured and unstructured data, and EHR structured and unstructured data with publicly available data). First, we assessed several machine learning algorithms—including XGBoost, 38 Random Forest, 36 Support Vector Machine (SVM), 39 and K-Nearest Neighbors (KNN) 40 methods in the structured EHR data. Performance comparisons were conducted to determine the optimal modeling method, which was subsequently applied to all remaining experiments across both study sites. Second, to process unstructured EHR data, multiple ML methods were used for aggregating individual document-level embeddings into patient-last-encounter-level representations: max pooling, mean pooling, and dimensionality reduction techniques such as Principal Component Analysis (PCA) 41 and t-distributed Stochastic Neighbor Embedding (t-SNE). 42 Features were eventually added to the ML model that was determined to have the highest predictive performance, measured by the best F1-score and AUC-ROC, on the structured EHR data in the first step. Thirdly, we evaluated whether the inclusion of cause of death information extracted from publicly available data using language models as described above would improve the performance of the selected ML modeling method using the pre-specified structured data feature management for both structured and unstructured EHR data. Statistical Analysis Site specific models were initially assessed at both institutions. The dataset was split into 3 sets: 10% for hyperparameter tuning, 70% for training, and 20% for testing. Model performance was evaluated using the weighted area under the receiver operating characteristic curve (AUC) and the weighted F-measure. AUC 43 incorporates class weights to address dataset imbalance, and the weighted F-measure balances precision and recall across all classes while considering class frequencies. 44 , 45 All models were trained and evaluated using causes of death derived from official death records— specifically, the NDI for VUMC and state health department death records for MGB—as the gold-standard reference. Correlation analyses compared CoD categories derived from public data with those from official death records. Spearman’s rank correlation coefficient and Kendall’s Tau assessed the strength and direction of associations between the two ranking sources. 46 A p-value threshold of <0.05 was used to determine statistical significance. We then conducted cross-institutional validation to assess model performance and generalizability across different healthcare institutions. Models trained on data from one institution (either VUMC or MGB) were tested on data from the other without any retraining or fine-tuning. This approach allowed us to evaluate how well the models generalized to new patient populations and clinical environments that differed from those of the training institution. Results Patient demographics and CoD distributions are summarized in Table 2 . The study included 13,708 deceased patients from VUMC and 34,839 from MGB. Malignant neoplasms were the leading CoD at both institutions, accounting for 4,155 deaths (30.3%) among VUMC patients and 10,961 deaths (31.5%) among MGB patients. Diseases of the heart were the second most common CoD, representing 2,192 deaths (16.0%) at VUMC and 6,445 deaths (18.5%) at MGB. COVID-19-related deaths comprised 1,044 cases (7.6%) at VUMC and 1,135 cases (3.3%) at MGB. View this table: View inline View popup Table 2. Distribution of Top 15 CoD* at VUMC and MGB Initial Model and Approach Selection At VUMC , the XGBoost model demonstrated superior performance, achieving a weighted AUC of 0.86 (95% CI, 0.82–0.88) ( Table 3 ). Random Forest achieved a weighted AUC of 0.79 (95% CI, 0.75–0.83). The SVM and KNN models exhibited lower performance, with weighted AUCs of 0.73 (95% CI, 0.70–0.76) and 0.65 (95% CI, 0.60–0.70), respectively. At MGB , the Random Forest model achieved the highest weighted AUC of 0.82 (95% CI, 0.82–0.83), slightly outperforming XGBoost, which had a weighted AUC of 0.80 (95% CI, 0.79–0.80). The SVM and KNN models showed lower performance at MGB as well, with weighted AUCs of 0.75 (95% CI, 0.74–0.75) and 0.61 (95% CI, 0.61–0.62), respectively. Based on the overall performance across both institutes , XGBoost was selected for subsequent analyses due to its superior performance at VUMC and competitive performance at MGB. View this table: View inline View popup Download powerpoint Table 3. Performance of Machine Learning Models Only Using Structured features at VUMC and MGB To enhance model performance by incorporating information from unstructured clinical notes, various methods for aggregating data features extracted at the document level into patient-level representations were assessed using the XGBoost model at both VUMC and MGB. At VUMC, max pooling of embeddings yielded the best performance, increasing the weighted AUC to 0.90 (95% CI, 0.88–0.93) ( Table 4 ). Other aggregation methods, including mean pooling, PCA, and t-SNE, also improved performance but to a lesser extent. At MGB, the mean pooling of embeddings resulted in the highest performance, with a weighted AUC of 0.92 (95% CI, 0.91–0.92). Max pooling also improved performance, achieving a weighted AUC of 0.88 (95% CI, 0.87–0.88). PCA and t-SNE methods showed lower performance at MGB, with weighted AUCs significantly lower than those achieved with pooling methods. Based on these results, max pooling was selected as the aggregation method at VUMC due to its superior performance, while mean pooling was selected at MGB. View this table: View inline View popup Download powerpoint Table 4. Performance of XGBoost Model including the features extracted using Unstructured Data Aggregation Methods at VUMC and MGB Model Performance: Structured Data Alone vs. Integrated Structured and Unstructured Data at Both Sites At both VUMC and MGB, integrating structured and unstructured data statistically improved model performance, as shown in Table 5 . At VUMC , the combined model achieved a weighted AUC of 0.90 (95% CI, 0.88–0.93) compared to 0.86 (95% CI, 0.84–0.87) with structured data alone ( Table 5 ). Similarly, at MGB , the combined model reached a weighted AUC of 0.92 (95% CI, 0.91–0.92), outperforming the structured-only model with a weighted AUC of 0.80 (95% CI, 0.79–0.80). View this table: View inline View popup Table 5. Model Performance with Structured Data Alone vs Combined Data at VUMC and MGB On an individual CoD level, as shown in Table 6 , the addition of unstructured data led to variable improvements across different categories in both sites. At VUMC, the incorporation of unstructured data yielded statistically significant improvements (based on non-overlapping confidence intervals) for several CoD: chronic lower respiratory disease (AUC: 0.69 [0.68-0.70] to 0.78 [0.77-0.79]), intentional self-harm (0.73 [0.71-0.75] to 0.82 [0.79-0.85]), and cerebrovascular disease (0.67 [0.65-0.69] to 0.75 [0.73-0.77]). Additionally, several other CoDs also demonstrated improvements with non-overlapping CIs. Changes in diabetes mellitus (0.82 [0.80-0.84] to 0.80 [0.78-0.82]) and heart diseases (0.99 [0.97-0.99] to 0.98 [0.97-0.99]) were not statistically significant. MGB demonstrated more substantial improvements, particularly in chronic lower respiratory disease (0.75 [0.71-0.77] to 0.91 [0.90-0.94]), diabetes mellitus (0.75 [0.70-0.77] to 0.93 [0.92-0.95]), Parkinson disease (0.63 [0.57-0.67] to 0.95[0.92-0.97]), and influenza/pneumonia (0.63 [0.61-0.70] to 0.93 [0.91-0.95]). All improvements at MGB were statistically significant based on non-overlapping confidence intervals. View this table: View inline View popup Table 6. AUC for Individual CoD at VUMC and MGB Public Data Integration Integration of public data across two major health systems revealed variable utility in CoD prediction modeling. At VUMC, linkage with public sources—including obituaries and memorial websites—identified 6,285 patients, of whom 810 (12.9%) had discernible CoD information; 526 cases mapped to the top 15 NCHS rankable categories. Similarly, at MGB, 2,298 patients were linked to public data, with 451 (19.6%) having identifiable CoD information, though only 137 cases corresponded to the top 15 NCHS categories. The incorporation of sixteen public data-derived CoD features at VUMC did not improve overall weighted AUC, and MGB showed no improvement, with decreases in predictive accuracy. Both institutions demonstrated strong rank correlations between public and healthcare datasets (VUMC: Spearman’s ρ=0.7785, Kendall’s τ=0.7113; MGB: Spearman’s ρ=0.6969, Kendall’s τ=0.5652; all p<0.05), suggesting alignment in CoD distributions despite limited utility for enhancing predictive models. Detailed analyses of ranking correlations and predictive performance following public data integration are provided in the supplementary materials. Cross-Site Model Evaluation When models trained at one institution were tested on data from the other, significant performance degradation was observed. The XGBoost model using structured data trained on VUMC data achieved a weighted AUC of 0.86 when tested on VUMC data but decreased when tested on MGB structured data ( Table 7 ). Similarly, the model trained on MGB data had a weighted AUC of 0.80 on MGB data but dropped when tested on VUMC data. Cross-site testing of VUMC-trained structured plus unstructured data models were also conducted and no substantial increase in cross-site performance was observed. View this table: View inline View popup Download powerpoint Table 7. Cross-Institutional Model Performance § Feature selection methods—including Stability Selection, Random Forest Selection, and SelectFromModel—were applied to enhance model interpretability and potentially improve cross-site performance. By reducing dimensionality to potentially address overfitting to a local environment, these methods improved the performance somewhat for certain causes of death. However, they did not substantially mitigate the performance degradation observed during cross-site evaluations uniformly across all causes ( Supplementary Table S2 ). Discussion In this multi-institutional retrospective cohort study, we developed and conducted both internal and external validations of machine learning models to predict CoD by integrating structured EHR data, unstructured clinical notes, and publicly available data sources. Utilizing data from VUMC and MGB, we established a comprehensive and scalable pipeline for feature extraction and model training that can be adapted by other healthcare institutions. Our findings demonstrate that integrating unstructured clinical data with structured EHR data enhances predictive performance within individual institutions and offers a framework for improved mortality surveillance. In addition, our findings support the need to locally tune or train the modeling pipeline to preserve performance. A key strength of our study is the development of a robust pipeline capable of processing large-scale data from diverse sources. By efficiently converting structured data—such as diagnoses, procedures, and laboratory results—and unstructured data—such as clinical notes—into machine learning–ready features, we enabled scalable modeling of patient information. The use of advanced NLP techniques, including strategies to aggregate information from the document level to the patient level, facilitated the integration of heterogeneous data types. Additionally, the utilization of standardized, high-quality gold-standard labels from official death records ensured the reliability and validity of our predictive models. Importantly, this pipeline was successfully deployed at both VUMC and MGB, allowing for local model training while maintaining methodological consistency. This approach is particularly beneficial for institutions with privacy concerns or varying data governance policies, as it permits site-specific adaptations without compromising data security. This work underscores the importance of moving beyond structured data features and attempting to incorporate unstructured data into predictive models. While unstructured information from free-text notes takes more effort in pre-processing compared to structured features alone, which are mostly readily available, the performance gain in prediction models can be substantial, as demonstrated at both institutions in our study. These improvements were particularly pronounced in specific cause-of-death categories, emphasizing the value of unstructured data in capturing nuanced clinical insights not easily available in structured formats. Additionally, we observed variation in model performance across CoD categories. Common and well-documented conditions (e.g., cancer, heart disease) had higher AUCs, while others (e.g., Alzheimer’s, self-harm) showed lower and more variable performance. This underscores the need for condition-specific validation and tailored model interpretation. Integrating clinical narratives represents a crucial improvement, as they frequently contain essential details that may provide additional clinical context to bolster predictive model accuracy. Despite the improvements observed within individual institutions, we encountered substantial performance degradation during cross-institutional validation. Models trained at VUMC performed inadequately when applied to MGB data, and vice versa. This decline underscores the significant impact of local data characteristics—such as differences in EHR systems, coding practices, and patient populations—on model performance. 47 , 48 To mitigate potential overfitting to institutional data, we evaluated various feature selection methods. These approaches improved the performance somewhat for certain causes of death but the cross-site performance drop was still observed. This finding reflects broader debates in the literature, where localized models are often preferred for their tailored accuracy, whereas global models offer broader but less precise generalizability. 48 , 49 Rather than attempting to directly transfer predictive models across institutions, which often leads to performance degradation due to these local variations, institutions may achieve better outcomes by adopting a shared, robust modeling process. 49 Such an approach allows for the development of models tailored to specific data while maintaining methodological consistency. By disseminating a standardized process, as demonstrated in this study, institutions can build more accurate and context-specific models, which aligns with prior research advocating for adaptable, locally optimized models over a one-size-fits-all solution. 47 , 50 , 51 The integration of publicly available data sources, such as obituaries and memorial websites, did not significantly improve overall model performance. This likely represents both a variability in how patients and their human networks publicly share information but also reflects a low rate of matching due to limitations in patient identifiers. These data sources have been thought to be valuable supplementary sources for timely mortality surveillance, particularly in situations where official death records are delayed or incomplete. However, while additional studies are needed to further evaluate the capacity of these data for this purpose, in this two-institution study we were unable to find significant improvements through incorporation of these data. Our study contributes to the growing body of literature emphasizing the importance of integrating structured and unstructured data in predictive modeling. The scalable and adaptable pipeline we developed offers a framework that can be implemented by other medical institutions to develop high-performing models tailored to local patient populations. This is particularly relevant for CoD prediction, where timely and accurate information is crucial for public health planning and resource allocation. Limitations This study has several limitations that should be considered when interpreting the findings. First, the heterogeneity of data between institutions—due to differences in EHR systems, coding practices, and patient demographics—may have contributed to the poor cross-site model performance. Models trained on data from one institution did not generalize well to another, indicating that local factors significantly influence model performance. This lack of generalizability underscores the need for site-specific model calibration to optimize predictive accuracy, which is critical for effective deployment across diverse healthcare environments. Second, the retrospective nature of data collection may introduce biases related to data availability and documentation practices during the study period. Variations in how data were recorded, missing information and differences in clinical documentation could affect the consistency and reliability of the input features used for model training and validation. In this study, the use of state death data for validation at MGB versus the NDI at VUMC is a limitation that may have impacted model comparability. This variation in death data sources introduces potential inconsistencies, which could influence the accuracy of CoD labeling and model generalizability across sites. Future Directions To address the challenges of generalizability and enhance the applicability of predictive models across different healthcare settings, several strategies warrant exploration. Federated learning presents a promising approach by enabling models to be trained collaboratively across multiple institutions without sharing sensitive patient data. 52 This method allows for broader learning from diverse datasets while maintaining patient privacy, potentially leading to more robust models that generalize better across institutions. Transfer learning and domain adaptation techniques also offer potential benefits. By leveraging knowledge gained from diverse datasets and allowing models to adjust to local variations, these techniques can improve model performance in new settings. 53 , 54 Advanced LLMs capable of handling both structured and unstructured data through few-shot learning represent another promising avenue. A unified LLM could streamline data integration and enhance generalizability by capturing complex insights across diverse data types, reducing the need for complex data-specific pipelines. 55 Additionally, cross-institutional training and continuous model updates with real-time data could enable models to adapt to evolving healthcare environments. This dynamic updating could improve long-term performance and ensure that models remain relevant as clinical practices and patient populations change over time. 56 Conclusion Our study demonstrates that integrating structured and unstructured EHR data enhances the predictive performance of machine learning models for CoD prediction within individual institutions. The development of a scalable and adaptable pipeline facilitates local model training and may improve mortality surveillance efforts. However, significant challenges remain in achieving cross-institutional generalizability, underscoring the need for site-specific tuning or alternative strategies to address local data heterogeneity. Funding statement This project was supported by Task Order 75F40123F19010 under Master Agreement 75F40119D10037 from the US Food and Drug Administration (FDA). FDA coauthors reviewed the study protocol, statistical analysis plan, and the manuscript for scientific accuracy and clarity of presentation. Representatives of the FDA reviewed a draft of the manuscript for the presence of confidential information and accuracy regarding the statement of any FDA policy. The views expressed are those of the authors and not necessarily those of the US FDA. Data Availability Statement The data used in this study are derived from electronic health records maintained by Vanderbilt University Medical Center (VUMC) and Massachusetts General Brigham (MGB), as well as mortality datasets provided by the National Death Index and the state health departments of Massachusetts, Connecticut, and Vermont. Access to these data is restricted due to institutional data use policies and patient privacy regulations. Code Availability The code can be downloaded at the following link: https://github.com/CPHI-TVHS/Enhancing-Cause-of-Death-Prediction S1: Supplementary: Public Data Integration Results View this table: View inline View popup Download powerpoint Table S 1. Overall weighted AUC AND AUC for Individual CoD at VUMC (S=Structured data S+U+PD= Structured +Unstructured +Public data) Comparative Analysis of Public Data Versus NDI/State Data on Causes of Death Download figure Open in new tab Figure s1: Rank Correlation of Top 15 Causes of Death - Public Data vs NDI Data. (NCHS_15 (Malignant Neoplasm), NCHS_23 (Diseases of Heart), NCHS_52 (COVID-19), NCHS_45 (Unintentional Injuries), NCHS_25 (Cerebrovascular Disease), NCHS_36 (Chronic Liver Disease and Cirrhosis), NCHS_30 (Chronic Lower Respiratory Disease), NCHS_18 (Diabetes Mellitus), NCHS_38 (Nephritis, Nephrotic Syndrome, and Nephrosis), NCHS_28 (Influenza and Pneumonia), NCHS_21 (Parkinson’s Disease), NCHS_24 (Essential Hypertension and Hypertensive Renal Disease), NCHS_22 (Alzheimer’s Disease), and Other (Other Causes)). Graph Overview Figure A (VUMC): The scatter plot illustrates a strong positive correlation between the ranking of causes of death in public and healthcare data, as shown by points closely following a diagonal trend line. This alignment indicates that more frequently mentioned causes of death in public data, such as Alzheimer’s Disease and essential hypertension, also receive higher rankings in National Death Index (NDI) data. A few deviations from this trend, like Alzheimer’s Disease (NCHS_22) and essential hypertension (NCHS_24), show higher ranks in public data compared to healthcare records. Figure B (MGB): Similar to VUMC, the scatter plot for MGB shows a positive correlation between the ranks in public and healthcare data, suggesting alignment in cause of death distributions across data sources. However, there are minor deviations from the trend line, such as certain conditions ranking higher in state data than in social media mentions. Together, Figures A and B suggest a strong rank correlation between public and institutional datasets across both VUMC and MGB, highlighting the consistency of cause of death distributions between public and healthcare records, with slight variations for specific conditions in each system. Spearman and Kendall Tau Correlations VUMC: The correlation coefficients between public and healthcare data are high, with Spearman’s ρ = 0.7785 (p = 0.0004) and Kendall’s τ = 0.7113 (p = 0.0001), indicating statistically significant correlations. These values suggest a strong, consistent relationship between public reporting and healthcare data for CoD, implying alignment in cause-specific prevalence across sources. MGB: Similarly, MGB demonstrates strong correlations, though slightly lower than VUMC, with Spearman’s ρ = 0.6969 (p = 0.0056) and Kendall’s τ = 0.5652 (p = 0.0058). These significant correlations reflect alignment between public perception and healthcare data at MGB, albeit with some variations in specific CoD rankings. S2: Supplementary: Feature selection results Feature selection methods—including Stability Selection, Random Forest Selection, and SelectFromModel—were applied to enhance model interpretability and potentially improve cross-site performance. Despite reducing dimensionality and enhancing interpretability, these methods did not substantially mitigate the performance degradation observed during cross-site evaluations. View this table: View inline View popup Table S2 Feature Selection Methods and Performance Footnotes Funding sources: The contents are those of the author(s) and do not necessarily represent the official views of, nor an endorsement, by FDA/HHS, or the U.S. Government. This project was supported by Task Order 75F40119F19002 under Master Agreement 75F40119D10037 from the US Food and Drug Administration (FDA). updated authors name and affiliation References 1. ↵ Weiss NS . All-cause mortality as an outcome in epidemiologic studies: proceed with caution . European journal of epidemiology . Mar 2014 ; 29 ( 3 ): 147 – 9 . doi: 10.1007/s10654-014-9899-y OpenUrl CrossRef PubMed 2. Flegal KM , Kit BK , Orpana H , Graubard BI . Association of All-Cause Mortality With Overweight and Obesity Using Standard Body Mass Index Categories: A Systematic Review and Meta-analysis . JAMA . 2013 ; 309 ( 1 ): 71 – 82 . doi: 10.1001/jama.2012.113905 OpenUrl CrossRef PubMed Web of Science 3. Berrington de Gonzalez A , Hartge P , Cerhan JR , et al. Body-Mass Index and Mortality among 1.46 Million White Adults . New England Journal of Medicine . 2010 ; 363 ( 23 ): 2211 – 2219 . doi: 10.1056/NEJMoa1000367 OpenUrl CrossRef PubMed Web of Science 4. ↵ Starfield B . Is US Health Really the Best in the World? JAMA . 2000 ; 284 ( 4 ): 483 – 485 . doi: 10.1001/jama.284.4.483 OpenUrl CrossRef PubMed Web of Science 5. ↵ Cowper DC , Kubal JD , Maynard C , Hynes DM . A Primer and Comparative Review of Major U.S. Mortality Databases . Annals of Epidemiology . 2002/10/01/ 2002 ; 12 ( 7 ): 462 – 468 . doi: 10.1016/S1047-2797(01)00285-X OpenUrl CrossRef PubMed Web of Science 6. ↵ Blackstone EH . Demise of a vital resource . The Journal of thoracic and cardiovascular surgery . Jan 2012 ; 143 ( 1 ): 37 – 8 . doi: 10.1016/j.jtcvs.2011.11.028 OpenUrl CrossRef PubMed 7. ↵ Auffray C , Balling R , Barroso I , et al. Making sense of big data in health research: towards an EU action plan . Genome medicine . 2016 ; 8 : 1 – 13 . OpenUrl PubMed 8. ↵ Jeblee S , Gomes M , Jha P , Rudzicz F , Hirst G . Automatically determining cause of death from verbal autopsy narratives . BMC medical informatics and decision making . 2019 ; 19 : 1 – 13 . OpenUrl CrossRef 9. ↵ Kim C , You SC , Reps JM , Cheong JY , Park RW . Machine-learning model to predict the cause of death using a stacking ensemble method for observational data . Journal of the American Medical Informatics Association . 2021 ; 28 ( 6 ): 1098 – 1107 . OpenUrl PubMed 10. Mapundu MT , Kabudula CW , Musenge E , Olago V , Celik T . Performance evaluation of machine learning and Computer Coded Verbal Autopsy (CCVA) algorithms for cause of death determination: A comparative analysis of data from rural South Africa . Frontiers in Public Health . 2022 ; 10 : 990838 . OpenUrl PubMed 11. Nunes PV , Mancine L , Neves BA , et al. Causes of death in individuals with lifetime major depression: a comprehensive machine learning analysis from a community-based autopsy center . BMC psychiatry . 2024 ; 24 ( 1 ): 531 . OpenUrl PubMed 12. Aggrawal R , Pal S . Sequential feature selection and machine learning algorithm-based patient’s death events prediction and diagnosis in heart disease . SN Computer Science . 2020 ; 1 ( 6 ): 344 . OpenUrl 13. ↵ Falissard L , Morgand C , Roussel S , et al. A deep artificial neural network− based model for prediction of underlying cause of death from death certificates: algorithm development and validation . JMIR Medical informatics . 2020 ; 8 ( 4 ): e17125 . OpenUrl 14. ↵ Kormilitzin A , Vaci N , Liu Q , Nevado-Holgado A . Med7: A transferable clinical natural language processing model for electronic health records . Artificial Intelligence in Medicine . 2021 ; 118 : 102086 . OpenUrl CrossRef PubMed 15. Bhatia A , Titus R , Porto JG , et al. Application of Natural Language Processing in Electronic Health Record Data Extraction for Navigating Prostate Cancer Care: A Narrative Review . Journal of Endourology . 2024 ; 16. Savova GK , Danciu I , Alamudun F , et al. Use of natural language processing to extract clinical cancer phenotypes from electronic medical records . Cancer research . 2019 ; 79 ( 21 ): 5463 – 5470 . OpenUrl Abstract / FREE Full Text 17. ↵ Yim W-w , Yetisgen M , Harris WP , Kwan SW. Natural language processing in oncology: a review . JAMA oncology . 2016 ; 2 ( 6 ): 797 – 804 . OpenUrl PubMed 18. ↵ Tourassi G , Yoon H-J , Xu S . A novel web informatics approach for automated surveillance of cancer mortality trends . Journal of biomedical informatics . 2016 ; 61 : 110 – 118 . OpenUrl CrossRef PubMed 19. Olayinka OO , Bayleyegn TM , Noe RS , Lewis LS , Arrisi V , Wolkin AF . Evaluation of real-time mortality surveillance based on media reports . Disaster medicine and public health preparedness . 2017 ; 11 ( 4 ): 460 – 466 . OpenUrl PubMed 20. ↵ Simonsen L , Gog JR , Olson D , Viboud C . Infectious disease surveillance in the big data era: towards faster and locally relevant systems . The Journal of infectious diseases . 2016 ; 214 ( suppl_4 ): S380 – S385 . OpenUrl CrossRef PubMed 21. ↵ Thirunavukarasu AJ , Ting DSJ , Elangovan K , Gutierrez L , Tan TF , Ting DSW . Large language models in medicine . Nature medicine . 2023 ; 29 ( 8 ): 1930 – 1940 . OpenUrl CrossRef PubMed 22. ↵ Singhal K , Azizi S , Tu T , et al. Large language models encode clinical knowledge . Nature . 2023 ; 620 (7972): 172 – 180 . OpenUrl CrossRef PubMed 23. ↵ Jiang LY , Liu XC , Nejatian NP , et al. Health system-scale language models are all-purpose prediction engines . Nature . 2023 ; 619 ( 7969 ): 357 – 362 . OpenUrl PubMed 24. ↵ Clusmann J , Kolbinger FR , Muti HS , et al. The future landscape of large language models in medicine . Communications medicine . 2023 ; 3 ( 1 ): 141 . OpenUrl PubMed 25. ↵ Statistics NCfH . ICD-10 Cause-of-Death Lists for Tabulating Mortality Statistics (Updated October 2020 to include WHO updates to ICD-10 for data year 2020) . NCHS instruction manual, part . 2020 ; 9 26. ↵ Skopp NA , Smolenski DJ , Schwesinger DA , Johnson CJ , Metzger-Abamukong MJ , Reger MA . Evaluation of a methodology to validate National Death Index retrieval results among a cohort of US service members . Annals of epidemiology . 2017 ; 27 ( 6 ): 397 – 400 . OpenUrl PubMed 27. ↵ Jamal-Allial A , Sponholtz T , Vojjala SK , et al. Validation of Mortality Data Sources Compared to the National Death Index in the Healthcare Integrated Research Database . Pragmatic and Observational Research . 2025 : 19 – 25 . 28. ↵ Wu P , Gifford A , Meng X , et al. Mapping ICD-10 and ICD-10-CM codes to phecodes: workflow development and initial evaluation . JMIR medical informatics . 2019 ; 7 ( 4 ): e14325 . OpenUrl CrossRef 29. ↵ Huff SM , Rocha RA , McDonald CJ , et al. Development of the logical observation identifier names and codes (LOINC) vocabulary . Journal of the American Medical Informatics Association . 1998 ; 5 ( 3 ): 276 – 292 . OpenUrl CrossRef PubMed 30. ↵ Nelson SJ , Zeng K , Kilbourne J , Powell T , Moore R . Normalized names for clinical drugs: RxNorm at 6 years . Journal of the American Medical Informatics Association . 2011 ; 18 ( 4 ): 441 – 448 . OpenUrl CrossRef PubMed 31. ↵ Li Y , Wehbe RM , Ahmad FS , Wang H , Luo Y . A comparative study of pretrained language models for long clinical text . Journal of the American Medical Informatics Association . 2023 ; 30 ( 2 ): 340 – 347 . OpenUrl CrossRef PubMed 32. ↵ Al-Garadi M , LeNoue-Newton M , Matheny ME , et al. Automated Extraction of Mortality Information from Publicly Available Sources Using Language Models . medRxiv . 2024 :2024.10. 28.24316027. 33. ↵ LeNoue-Newton M , Al-Garadi MA , Ngan K , et al. Augmenting Fact and Date of Death in Electronic Health Records using Internet Media Sources: A Validation Study from Two Large Healthcare Systems . medRxiv . 2025 :2025.01. 24.25321042. 34. ↵ Meinshausen N , Bühlmann P . Stability selection . Journal of the Royal Statistical Society Series B: Statistical Methodology . 2010 ; 72 ( 4 ): 417 – 473 . OpenUrl CrossRef Web of Science 35. ↵ Ranstam J , Cook JA . LASSO regression . Journal of British Surgery . 2018 ; 105 ( 10 ): 1348 – 1348 . OpenUrl CrossRef 36. ↵ Breiman L . Random forests . Machine learning . 2001 ; 45 : 5 – 32 . OpenUrl CrossRef 37. ↵ Badrouchi S , Ahmed A , Bacha MM , Abderrahim E , Abdallah TB . A machine learning framework for predicting long-term graft survival after kidney transplantation . Expert Systems with Applications . 2021 ; 182 : 115235 . OpenUrl 38. ↵ Chen T , Guestrin C. Xgboost: A scalable tree boosting system . 2016 : 785 – 794 . 39. ↵ Hearst MA , Dumais ST , Osuna E , Platt J , Scholkopf B . Support vector machines . IEEE Intelligent Systems and their applications . 1998 ; 13 ( 4 ): 18 – 28 . OpenUrl CrossRef 40. ↵ Zhang Z . Introduction to machine learning: k-nearest neighbors . Annals of translational medicine . 2016 ; 4 ( 11 ) 41. ↵ Abdi H , Williams LJ . Principal component analysis . Wiley interdisciplinary reviews: computational statistics . 2010 ; 2 ( 4 ): 433 – 459 . OpenUrl 42. ↵ Hinton GE , Roweis S . Stochastic neighbor embedding . Advances in neural information processing systems . 2002 ; 15 43. ↵ Bradley AP . The use of the area under the ROC curve in the evaluation of machine learning algorithms . Pattern recognition . 1997 ; 30 ( 7 ): 1145 – 1159 . OpenUrl CrossRef Web of Science 44. ↵ Powers DM . Evaluation: from precision, recall and F-measure to ROC, informedness, markedness and correlation . arXiv preprint arXiv:201016061 . 2020 ; 45. ↵ Keilwagen J , Grosse I , Grau J . Area under precision-recall curves for weighted and unweighted data . PloS one . 2014 ; 9 ( 3 ): e92209 . OpenUrl CrossRef PubMed 46. ↵ Puth M-T , Neuhäuser M , Ruxton GD . Effective use of Spearman’s and Kendall’s correlation coefficients for association between two measured traits . Animal Behaviour . 2015 ; 102 : 77 – 84 . OpenUrl 47. ↵ Amrollahi F , Shashikumar SP , Holder AL , Nemati S . Leveraging clinical data across healthcare institutions for continual learning of predictive risk models . Scientific reports . 2022 ; 12 ( 1 ): 8380 . OpenUrl PubMed 48. ↵ Curth A , Thoral P , van den Wildenberg W , et al. Transferring clinical prediction models across hospitals and electronic health record systems . Springer ; 2020 : 605 – 621 . 49. ↵ Barak-Corren Y , Castro VM , Nock MK , et al. Validation of an electronic health record–based suicide risk prediction modeling approach across multiple health care systems . JAMA network open . 2020 ; 3 ( 3 ): e201262 – e201262 . OpenUrl 50. ↵ Kopitar L , Cilar L , Kocbek P , Stiglic G . Local vs. global interpretability of machine learning models in type 2 diabetes mellitus screening . Springer ; 2019 : 108 – 119 . 51. ↵ Carballa Smichowski B. Alternative data governance models: Moving beyond one-size-fits-all solutions . Intereconomics . 2019 ; 54 ( 4 ): 222 – 227 . OpenUrl 52. ↵ Antunes RS , André da Costa C , Küderle A , Yari IA , Eskofier B . Federated learning for healthcare: Systematic review and architecture proposal . ACM Transactions on Intelligent Systems and Technology (TIST) . 2022 ; 13 ( 4 ): 1 – 23 . OpenUrl 53. ↵ Gu T , Lee PH , Duan R . COMMUTE: communication-efficient transfer learning for multi-site risk prediction . Journal of biomedical informatics . 2023 ; 137 : 104243 . OpenUrl PubMed 54. ↵ Li S , Cai T , Duan R . Targeting underrepresented populations in precision medicine: A federated transfer learning approach . The Annals of Applied Statistics . 2023 ; 17 ( 4 ): 2970 – 2992 . OpenUrl PubMed 55. ↵ Guo LL , Fries J , Steinberg E , et al. A multi-center study on the adaptability of a shared foundation model for electronic health records . NPJ Digital Medicine . 2024 ; 7 ( 1 ): 171 . OpenUrl PubMed 56. ↵ Jenkins DA , Martin GP , Sperrin M , et al. Continual updating and monitoring of clinical prediction models: time for dynamic prediction systems? Diagnostic and Prognostic Research . 2021 ; 5 : 1 – 7 . OpenUrl View the discussion thread. Back to top Previous Next Posted June 27, 2025. Download PDF Data/Code Email Thank you for your interest in spreading the word about medRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following Enhancing Cause of Death Prediction: Development and Validation of ML Models Using Multimodal Data Across Multiple Healthcare Sites Message Subject (Your Name) has forwarded a page to you from medRxiv Message Body (Your Name) thought you would like to see this page from the medRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share Enhancing Cause of Death Prediction: Development and Validation of ML Models Using Multimodal Data Across Multiple Healthcare Sites Mohammed Al-Garadi , Rishi J Desai , Kerry Ngan , Michele LeNoue-Newton , Ruth M. Reeves , Daniel Park , Jose J. Hernández-Muñoz , Shirley V. Wang , Judith C. Maro , Candace C. Fuller , Joshua Lin Kueiyu , Aida Kuzucan , Kevin Coughlin , Haritha Pillai , Melissa McPheeters , Jill Whitaker , Jessica A. Deere , Michael F. McLemore , Dax M. Westerman , Tony Morrow , Margaret A. Adgent , Michael E. Matheny medRxiv 2025.06.24.25330213; doi: https://doi.org/10.1101/2025.06.24.25330213 Share This Article: Copy Citation Tools Enhancing Cause of Death Prediction: Development and Validation of ML Models Using Multimodal Data Across Multiple Healthcare Sites Mohammed Al-Garadi , Rishi J Desai , Kerry Ngan , Michele LeNoue-Newton , Ruth M. Reeves , Daniel Park , Jose J. Hernández-Muñoz , Shirley V. Wang , Judith C. Maro , Candace C. Fuller , Joshua Lin Kueiyu , Aida Kuzucan , Kevin Coughlin , Haritha Pillai , Melissa McPheeters , Jill Whitaker , Jessica A. Deere , Michael F. McLemore , Dax M. Westerman , Tony Morrow , Margaret A. Adgent , Michael E. Matheny medRxiv 2025.06.24.25330213; doi: https://doi.org/10.1101/2025.06.24.25330213 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Health Informatics Subject Areas All Articles Addiction Medicine (569) Allergy and Immunology (863) Anesthesia (300) Cardiovascular Medicine (4442) Dentistry and Oral Medicine (444) Dermatology (383) Emergency Medicine (609) Endocrinology (including Diabetes Mellitus and Metabolic Disease) (1511) Epidemiology (15230) Forensic Medicine (30) Gastroenterology (1126) Genetic and Genomic Medicine (6610) Geriatric Medicine (668) Health Economics (998) Health Informatics (4542) Health Policy (1370) Health Systems and Quality Improvement (1613) Hematology (543) HIV/AIDS (1266) Infectious Diseases (except HIV/AIDS) (15923) Intensive Care and Critical Care Medicine (1103) Medical Education (623) Medical Ethics (147) Nephrology (668) Neurology (6607) Nursing (346) Nutrition (999) Obstetrics and Gynecology (1146) Occupational and Environmental Health (957) Oncology (3338) Ophthalmology (974) Orthopedics (369) Otolaryngology (420) Pain Medicine (436) Palliative Medicine (130) Pathology (665) Pediatrics (1693) Pharmacology and Therapeutics (692) Primary Care Research (712) Psychiatry and Clinical Psychology (5448) Public and Global Health (9239) Radiology and Imaging (2202) Rehabilitation Medicine and Physical Therapy (1370) Respiratory Medicine (1196) Rheumatology (596) Sexual and Reproductive Health (714) Sports Medicine (530) Surgery (712) Toxicology (99) Transplantation (289) Urology (265) (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'a01f2d62accedfa9',t:'MTc3OTgyMzY0Nw=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();