Survival Machine Learning Methods for Mortality Prediction After Heart Transplantation

doi:10.21203/rs.3.rs-2968288/v1

Download PDF

Article

Survival Machine Learning Methods for Mortality Prediction After Heart Transplantation

https://doi.org/10.21203/rs.3.rs-2968288/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Although prognostic models for heart transplantation (HTx) have been developed, a comprehensive benchmarking of survival machine learning methods for mortality prognosis has not been performed. Futhermore, assessing mortality in the most contemporary era of heart transplants following the 2018 donor heart allocation policy change is warranted. This study included 7,160 adult heart-only transplant recipients in the Scientific Registry of Transplant Recipients database who received their first transplant on or after October 18, 2018, with at least one recorded follow-up visit until June 3, 2021. Each model was trained using a nested 5-repeat, 5-fold cross-validation. Seven statistical and machine learning algorithms – Lasso, Ridge, Elastic Net, Gradient Boost, Extreme Gradient Boost Linear, Extreme Gradient Boost Tree, and Random Survival Forests were benchmarked against Cox PH in both post-policy and pre-policy eras. In the post-policy era, there was higher discriminatory power of machine learning models for one-year all-cause mortality compared to Cox. Bilirubin, age, BMI, total ischemic time, donor age, and history of ischemic cardiomyopathy were some of the most predictive selected variables. ECMO was highly predictive in the pre-policy cohort, but less so in the post-policy cohort. Machine learning (particularly tree-based ensemble) survival models can be used to capture complex short-term posttransplant mortality outcomes.

Health sciences/Cardiology

Health sciences/Risk factors

Biological sciences/Computational biology and bioinformatics/Machine learning

heart transplantation

mechanical circulatory support

mortality

machine learning

Orthotopic heart transplantation (HTx) remains the gold standard therapy for end-stage heart failure with a median survival of 12.5 years and an additional 2 years having survived one year post-transplant (1). Accurately predicting one-year mortality can aid clinical decision-making for posttransplant survival after HTx. Although many factors associated with post-transplantation survival have been identified (2, 3), more complex relationships among these variables have not been extensively explored. Even though previous studies have applied machine learning algorithms to predict heart transplant mortality with varying degrees of success (4–6), they use binary classification instead of survival analysis methods. Ignoring time-to-event information is known to result in less powerful and less comprehensive prediction estimates (7). Furthermore, posttransplantation survival is known to be non-proportional across different strata, making conventional statistical methods such as Cox proportional hazards regression potentially invalid and less powerful for assessing post-transplantation mortality among heart transplant recipients.

Another potentially important consideration for assessing one-year mortality prognosis is the revision of the donor heart allocation system on October 18, 2018 to “reduce wait-list mortality, enhance geographic organ sharing, and improve organ distribution equity” (8). Although it has been shown that 1-year posttransplant survival under the new heart allocation policy was not significantly different from before (9), it iss possible that both the clinical profile of transplant recipients and clinical care practices have shifted in the almost 5 years since policy implementation (10). Thus, in this study, we explore and benchmark the performance of machine learning survival algorithms in the post-2018 transplant era to better understand prognostic factors associated with one-year mortality. We compare these results to a parallel analysis run in a pre-2018 transplant era cohort.

Data Source

We used data from the Scientific Registry of Transplant Recipients (SRTR). The SRTR data system includes data on all donor, waitlisted candidates, and transplant recipients in the US, submitted by the Organ Procurement and Transplantation Network (OPTN) members. The Health Resources and Services Administration (HRSA), U.S. Department of Health and Human Services, provides oversight of the activities of the OPTN and SRTR contractors.

Study Population

We excluded recipients < 18 years of age at transplant (n = 1,275) and an additional 8,187 recipients of multi-organ transplants because allocation criteria are different for multi-organ and pediatric candidates. This study includes 7,160 adult heart-only transplant recipients who received their first transplant on or after October 18, 2018, with at least one recorded follow-up visit. We had follow-up data up until June 3, 2021. This study was reviewed and approved by the Institutional Review Board at the Harvard T.H. Chan School of Public Health.

Data Preparation

The outcome of interest was one-year all-cause mortality assessed from the time of heart transplantation to death or end of the one-year follow-up. Survival data for recipients in both cohorts were administratively censored at 1 year after transplant to prevent bias from the differential length of follow-up between cohorts. We excluded potential predictors with over 20% missingness. For those with less than 20% missingness, we used multiple imputation with chained equations using the R package mice (11), which preserves the associations in the data and the uncertainty in those associations. We assumed that the data are missing at random and that censoring is non-informative. We used one-hot encoding to to convert categorical variables into binary variables for each level of each categorical feature. Certain variables that had many levels with only a small number of observations were combined (Supplementary Appendix). All continuous features were normalized using a z-score transformation. This data pre-processing led to a final set of 80 demographic, clinical, recipient, waitlist, donor, and procedural variables (119 after one-hot encoding), of which 9 were categorical, 45 binary, and the remaining 26 numeric (Supplementary Table 1).

We also prepared a seasonally matched pre-policy cohort of patients transplants from November 1, 2014 with follow-up time to June 3, 2017 as a sensitivity analysis to account for known seasonal trends in decreased donor heart donation.(12) We also selected a pre-policy cohor thtat ended 1 year before policy implementation to avoid conoufnding from any anticipatory practice changes before the policy implementation in October 2018.

Machine Learning Analysis

We compared a Cox proportional hazards model and 7 machine learning algorithms. Lasso, Elastic Net, and Ridge regression were selected as the penalized cox regression methods (13, 14). Boosted cox regression methods included a Cox model with gradient boosting (Gradient Boost) (15) and Extreme Gradient Boosting (XGBoost) (16) with linear model-based boosting (XGBL) and tree-based boosting XGBoost (XGBT) (17). Random Survival Forests (RSF) (18) was chosen as the tree-based ensemble method. Details on the models and selected parameters are described in the Supplementary Appendix. Each model was trained using a nested 5-repeat, 5-fold cross-validation, and tuning was automatically performed for the machine learning algorithms to select optimal hyperparameters. A random search with 25 iterations was used to select values for the hyperparameters in the inner loop and model performance was evaluated in the outer loop. All model selection steps were standardized for every pair of training/test data. The R package mlr (Machine Learning in R) (19) was used as a framework to carry out benchmarking of these experiments. To evaluate model performance, we used the cross-validated concordance index (Harrell’s C-Index), which measures the proportion of pairs where the observation with the higher survival time has the higher probability of survival as predicted by the model. Variable weights were computed as the beta coefficients from regression-based models and importance from boosted and ensemble models. Overall variable relevance was taken simply as the number of times it was selected across 25 iterations. All analyses were performed using R 4.3.0. Code used in the analysis can be found at https://github.com/latlio/srtr_mlr.

Cohort Characteristics

Overall, 586 patients died within one year of their heart transplant during the study period from October 18, 2018, to June 3, 2021. The crude one-year survival rate was 90.9% (95% CI: 90.2%, 91.7%) with 5864 total person-years of follow-up time. The average recipient age was 53.5, and the average donor age was 32.5. A total of 72.5% (n = 5193) of the cohort was male, and 72.1% (n = 5162) of the cohort was white. 27.1% had Diabetes Mellitus. In this post-2018 cohort, 37.6% of patients were bridged with ventricular assist device (VAD), 4.9% of patients were bridged with extracorporeal membrane oxygenation (ECMO), 26.8% of patients received intra-aortic balloon pump therapy (IABP), patients spent a median of 33 days on the wait-list, and patients experienced 3.4 hours of ischemia time on average during transplant. A comprehensive summary of demographic, clinical, recipient, waitlist, donor, and procedural variables is detailed in Table 1.

Table 1

Baseline characteristics of heart transplant recipients from October 18, 2018 to June 3, 2021
	Overall (N = 7160)
Allocation Status
Previous 3-tier system	27 (0.4%)
2018 Status 1	587 (8.2%)
2018 Status 2	3335 (46.6%)
2018 Status 3	1396 (19.5%)
2018 Status 4	1427 (19.9%)
2018 Status 5	0 (0%)
2018 Status 6	388 (5.4%)
Demographic Variables
BMI (kg/m2)
Mean (SD)	27.8 (5.02)
Median [Min, Max]	27.5 [12.2, 47.5]
Highest Education Received
Below High School	243 (3.4%)
High School Graduate	2549 (35.6%)
College Graduate and beyond	4036 (56.4%)
Gender
Male	5193 (72.5%)
Female	1967 (27.5%)
Insurance
Medicaid	968 (13.5%)
Medicare	2372 (33.1%)
Private	3262 (45.6%)
Other	314 (4.4%)
Age
Mean (SD)	53.5 (12.9)
Median [Min, Max]	57.0 [18.0, 76.0]
Clinical History Variables
Cardiac Etiology
Valvular Heart Disease	74 (1.0%)
Dilated Myopathy	1302 (18.2%)
Ischemic Dilated Myopathy	1900 (26.5%)
Restrictive Myopathy	330 (4.6%)
Coronary Artery Disease	132 (1.8%)
Other	653 (9.1%)
Idiopathic	2701 (37.7%)
Dialysis
No	7049 (98.4%)
Yes	111 (1.6%)
Type 2 Diabetes
No	5174 (72.3%)
Yes	1938 (27.1%)
Previous Malignancy
No	7160 (100%)
Yes	0 (0%)
Prior Cardiac Surgery
No	5644 (78.8%)
Yes	1202 (16.8%)
Steroid Use
No	6441 (90.0%)
Yes	353 (4.9%)
Cytomegalovirus
Negative	0 (0%)
Positive	0 (0%)
Serum Creatinine (mg/dL)
Mean (SD)	1.21 (0.480)
Median [Min, Max]	1.13 [0.200, 8.00]
Epstein-Barr Virus
Negative	0 (0%)
Positive	0 (0%)
Functional Status At Time of Transplant
Moribund/hospitalized/severely disabled	4302 (60.1%)
Significant Assistance	1638 (22.9%)
Normal	518 (7.2%)
HBV Antibody
Negative	0 (0%)
Positive	0 (0%)
HBV Surf Antigen
Negative	0 (0%)
Positive	0 (0%)
Hepatitis C Virus
Negative	0 (0%)
Positive	0 (0%)
HIV Status
No	0 (0%)
Yes	0 (0%)
Candidate Variables
Time on Waitlist
Mean (SD)	186 (385)
Median [Min, Max]	33.0 [0, 5150]
Blood Type
A	2852 (39.8%)
AB	370 (5.2%)
B	1100 (15.4%)
O	2838 (39.6%)
Accept a Donor After Cardiac Death
No	6720 (93.9%)
Yes	440 (6.1%)
Accept a Hepatitis B Positive Donor
No	3167 (44.2%)
Yes	3993 (55.8%)
Accept a Hepatitis C Positive Donor
No	2401 (33.5%)
Yes	4759 (66.5%)
Accept a Donor with History of Heart Disease
No	2856 (39.9%)
Yes	3218 (44.9%)
Symptomatic Cerebrovascular Disease
No	6556 (91.6%)
Yes	512 (7.2%)
Candidate Functional Status
Moribund/hospitalized/severely disabled	3870 (54.1%)
Significant Assistance	2275 (31.8%)
Normal	663 (9.3%)
Maximum Age of Donor that Candidate Accepts
Mean (SD)	740 (125)
Median [Min, Max]	720 [65.0, 1190]
Maximum Distance of Donor that Candidate Accepts
Mean (SD)	1410 (995)
Median [Min, Max]	1250 [20.0, 10000]
Maximum Weight (kg) of Donor that Candidate Accepts
Mean (SD)	178 (49.2)
Median [Min, Max]	181 [59.0, 295]
Minimum Age of Donor that Candidate Accepts
Mean (SD)	119 (68.1)
Median [Min, Max]	144 [0, 1180]
Minimum Weight (kg) of Donor that Candidate Accepts
Mean (SD)	55.8 (14.1)
Median [Min, Max]	55.0 [15.0, 110]
Candidate Preliminary Cross-Matching Required
No	6642 (92.8%)
Yes	518 (7.2%)
Waitlist Glomerular Filtration Rate (MDRD)
Mean (SD)	68.7 (28.2)
Median [Min, Max]	64.2 [4.91, 522]
Recipient Variables
Glomerular Filtration Rate (MDRD) At Time of Transplant
Mean (SD)	70.1 (30.0)
Median [Min, Max]	65.1 [6.84, 494]
Number of HLA-A Mismatches
Mean (SD)	1.45 (0.612)
Median [Min, Max]	2.00 [0, 2.00]
Number of HLA-B Mismatches
Mean (SD)	1.70 (0.498)
Median [Min, Max]	2.00 [0, 2.00]
Number of HLA-DR Mismatches
Mean (SD)	1.50 (0.591)
Median [Min, Max]	2.00 [0, 2.00]
Received ECMO Pre-transplant
No	6810 (95.1%)
Yes	350 (4.9%)
Specific HLA Antibody
No	5674 (79.2%)
Yes	728 (10.2%)
HLA Typing Done
No	525 (7.3%)
Yes	6411 (89.5%)
Received IABP
No	5238 (73.2%)
Yes	1922 (26.8%)
Received Inotropes
No	4531 (63.3%)
Yes	2629 (36.7%)
Received Life Support (e.g. ECMO, IABP, PGE, Inotropes, Ventilator, NO, VAD)
No	1196 (16.7%)
Yes	5663 (79.1%)
Received Other Type of Life Support
No	6791 (94.8%)
Yes	369 (5.2%)
Medical Condition
ICU	3559 (49.7%)
Hospitalized Not in ICU	907 (12.7%)
Not Hospitalized	2389 (33.4%)
Number of HLA Mismatches
Mean (SD)	4.66 (1.06)
Median [Min, Max]	5.00 [0, 6.00]
Total Bilirubin (mg/dL)
Mean (SD)	1.01 (1.78)
Median [Min, Max]	0.700 [0.100, 51.0]
Received Transfusions
No	5861 (81.9%)
Yes	950 (13.3%)
Received Ventilator Support
No	5758 (80.4%)
Yes	1061 (14.8%)
VAD Usage
No	4232 (59.1%)
Yes	2695 (37.6%)
Donor Variables
Donor Blood Type
A	2557 (35.7%)
AB	134 (1.9%)
B	787 (11.0%)
O	3682 (51.4%)
Donor Age
Mean (SD)	32.5 (10.6)
Median [Min, Max]	32.0 [9.00, 70.0]
Donor Anti-CMV Antibody
Negative	0 (0%)
Positive	0 (0%)
Donor Anti-HCV Antibody
Negative	0 (0%)
Positive	0 (0%)
Donor Cause of Death
Anoxia	3202 (44.7%)
Cerebrovascular/Stroke	956 (13.4%)
Head Trauma	2795 (39.0%)
CNS Tumor	23 (0.3%)
Donor Cardiac Arrest
No	6327 (88.4%)
Yes	482 (6.7%)
Donor Serum Creatinine (mg/dL)
Mean (SD)	1.67 (1.76)
Median [Min, Max]	1.04 [0.0400, 17.9]
DDVAP Administered to Donor
No	6433 (89.8%)
Yes	552 (7.7%)
Donor Gender
Male	5132 (71.7%)
Female	2028 (28.3%)
Donor Height (cm)
Mean (SD)	174 (9.45)
Median [Min, Max]	175 [122, 206]
Donor Serum Creatinine > 1.5 mg/dL
No	5164 (72.1%)
Yes	1994 (27.8%)
Donor History of Cancer
No	7008 (97.9%)
Yes	66 (0.9%)
Donor Cigarette Use > 20 Pack-years
No	6149 (85.9%)
Yes	855 (11.9%)
Donor History of Cocaine Use
No	4981 (69.6%)
Yes	1875 (26.2%)
Donor History of Diabetes
No	278 (3.9%)
Yes	6816 (95.2%)
Donor History of Hypertension
No	1104 (15.4%)
Yes	5982 (83.5%)
Donor History of Drug Use
No	2530 (35.3%)
Yes	4372 (61.1%)
Donor History of Inotrope Use
No	4530 (63.3%)
Yes	2448 (34.2%)
Non-Beating Heart Donor
No	6968 (97.3%)
Yes	191 (2.7%)
Donor Weight (kg)
Mean (SD)	84.9 (20.2)
Median [Min, Max]	81.8 [39.9, 225]
Procedural Variables
Recipient Cardiac Output (L/min)
Mean (SD)	4.43 (1.43)
Median [Min, Max]	4.27 [0.200, 14.3]
Total Ischemic Time (min)
Mean (SD)	206 (64.8)
Median [Min, Max]	205 [20.0, 720]
Pulmonary Capillary Wedge Pressure (mmHg)
Mean (SD)	17.8 (8.73)
Median [Min, Max]	17.0 [0, 50.0]
Transplant Procedure Type
Orthotopic Bicaval	5677 (79.3%)
Orthotopic Traditional	977 (13.6%)
Orthotopic Total	188 (2.6%)
Heterotopic	0 (0%)
Diastolic Blood Pressure (mmHg)
Mean (SD)	19.3 (8.67)
Median [Min, Max]	18.0 [0, 110]
Mean Blood Pressure (mmHg)
Mean (SD)	27.0 (10.1)
Median [Min, Max]	26.0 [0, 110]
Systolic Blood Pressure (mmHg)
Mean (SD)	39.6 (13.8)
Median [Min, Max]	38.0 [0, 158]

Machine Learning Benchmark Performance

Machine learning algorithms performed slightly better than Cox with a 5-repeat 5-fold cross validation average C-index ranging from 0.614 to 0.509 (Fig. 1). Cox Boost (mean = 0.614 [SD = 0.029]), Ridge (mean = 0.608 [SD = 0.020]), and Random Survival Forests (mean = 0.608 [SD = 0.023]) were the strongest performers. Among the penalized Cox regression models, Ridge regression performed the best while Lasso performed the worst (mean = 0.51 [SD = 0.025]). Both XGBL (mean = 0.607 [SD = 0.026]) and XGBT (mean = 0.605 [SD = 0.021]) performed similarly. Generally, all the machine learning algorithms had a low standard deviation (range: 0.019–0.039) across cross-validation repeats, suggesting resilience to variations in the training data.

In the pre-2018 seasonally-matched cohort of 6,108 transplanted patients with 543 deaths, the machine learning models were generally less predictive with average C-indices ranging from 0.563 to 0.500 (Supplementary Fig. S1). We found that Lasso and Elastic Net struggled to have predictive power.

Prognostic Variables for Posttransplant One-year Mortality

Sparse models such as Lasso, Elastic Net, and Gradient Boost use a regularization term to shrink coefficients of variables to zero, effectively generating a subset of variables. Hereafter, we refer to this as a model’s “selection”, although we clarify that we did not perform explicit feature selection prior to training the model. Since we performed 5 repeats of 5-fold cross-validation, a variable can be selected 25 times at most. Variables selected in at least 20 of the repeated models across all sparse algorithms were bilirubin, age, BMI, total ischemic time, donor age, and history of ischemic cardiomyopathy (Fig. 2). Non-overlappingly, Lasso selected VAD, Medicaid insurance, dialysis, diabetes, and creatinine at least 20 times and Elastic Net selected a similar set with the addition of high school education and history of dilated cardiomyopathy (Supplementary Fig. S2). The Gradient Boost model selected a different subset of variables in all 25 iterations: blood transfusion, positive HBV Surf Antigen, unknown functional status, number of HLA-A mismatches, positive Anti-CMV Antigen, donor age, and donor O blood type. It is important to acknowledge the limitation that in situtations where predictor variables vary in their scale of measurement of their number of categories, the variables considered important may be artificially selected and highly variable (20). A comprehensive heatmap summary of the total number of times each input predictor was selected in each model is provided in Supplementary Fig. 3. In the pre-policy 2018 cohort, a different set of top variables were selected (Supplementary Fig. S4).

The strongest risk factors for one-year all-cause mortality amongst the penalized regression methods in terms of weight were generally age, BMI, diabetes, dialysis, creatinine, prior cardiac surgery, previous malignancies, bilirubin, VAD, transfusion, time spent on waitlist, waitlist status change, minimum accepted donor age, donor age, ischemic time, and baseline GFR (Supplementary Fig. S5-S7). Gradient Boost selected a different set of important variables, including transfusion, number of HLA mismatches, waitlist status change, pulmonary arterial pressures (Supplementary Fig. 8). XGBL and XGBT selected a similar set of variables as the penalized regression methods (Supplementary Fig. S9-S10). RSF prioritized age, bilirubin, candidate accepted characteristics (minimum and maximum age, minimum and maximum weight), donor height and weight, and pulmonary arterial pressures (Supplementary Fig. S11).

Some of the variables selected by RSF, the most predictive model in the pre-policy cohort were notably different, such as ECMO and maximum accepted distance for heart transplantation (Supplementary Fig. S12).

In this study, we comprehensively assessed survival machine learning methods for prognosis of one-year all-cause mortality in adult heart transplant recipients transplanted after the UNOS allocation policy change in October 2018 (8). In the post-policy era, there was higher discriminatory power of machine learning models for one-year all-cause mortality (range of mean C-index: 0.509–0.614) compared to Cox. Bilirubin, age, BMI, total ischemic time, donor age, and history of ischemic cardiomyopathy were some of the most predictive selected variables. ECMO was highly predictive in the pre-policy cohort, but less so in the post-policy cohort.

Across all algorithms that we assessed, RSF and Gradient Boost performed the best. Incorporating time-to-event information is preferable to logistic regression to build post-transplant mortality prediction models, as was done in previous studies (2, 4, 21, 22). In particular, Gradient Boost was previously found to be the strongest model in psychiatric data (23) and Random Survival Forests have been previously explored in heart transplant data (3, 24) and lung transplant data (25). These more complex models have an advantage over Cox regression in 2 aspects in that they do not require the proportional hazards assumption and and they can capture nonlinear interactions between outcome and variables.

Gradient Boost and RSF selected a different subset of predictive variables, but achieved similar levels of predictive performance. Gradient Boost identified previous malignancies (26), number of HLA mismatches (27), and receiving transfusions (28) as important predictors for one-year mortality, which have some published evidence. On the other hand, RSF selected flexibility with donor heart requirements (29) in addition to other well-established risk factors (24). It is important to note that these models are often more difficult to implement and interpret than penalized Cox regression methods since many hyperparameters can be tuned and there is no easily interpretable measure of directional association. In general, recipient characteristics tended to be more highly predictive compared to donor characteristics. This corroborates with findings in previous mortality modeling studies on earlier transplant eras (30, 31).

In our study, ECMO was not a strong predictor in the post-2018 cohort, although we found that it was in a pre-2018 cohort, which is corroborated in another study (3). This may reflect the increased usage of mechanical circulatory devices across all waitlist patients regardless of listed status (32), which limits these variables’ discriminative potential. One simulation study found that the increase in ECMO in the post-2018 cohort was statistically significant, which challenges the implicit assumption that the 2018 heart allocation policy would effectively prevent manipulation of waitlist priority via provision of mechanical support therapy (33). Furthermore, in the pre-policy 2018 we found that the maximum acceptable distance for a heart to travel for transplant was associated with lower mortality, which corroborates previous findings (34). We did not observe this level of importance for distance in the post-policy 2018 cohort which may suggest that other variables are now more predictive in an era where the new allocation policy has aimed to reduce waitlist times by extending donor distances (35).

Study Limitations

First, as with most national registries, SRTR is susceptible to data entry errors and lack of data regarding organ recovery or failure necessary to create continuously updated mortality estimates. However, the SRTR conducts edit checks, validation of data at time of entry, and internal verification when there are outliers. Second, the data did not include known prognostic risk factors such as natriuretic peptides, serum sodium, specific hemodynamic data such as heart rate, although these are usually available and collected at individual centers. In addition, SRTR does not include other variables such as patient adherence and granular clinical decision-making variables that likely affect survival. Notably, status justification form variables had a high proportion of missingness (above 90%), so they were not included in the predictor set. Third, this study is exploratory, and the intent was not to build a prediction model for individual prediction of mortality risk, but rather to identify potentially important variables associated with mortality. Fourth, we made the assumption that there was only missingness at random and non-informative censoring. However, there may be informative censoring, since early post-transplant deaths are more likely to be recorded than routine follow-up appointments in the first year after transplantation. We chose to study one-year mortality (which we administratively censored) given its clinical relevance for longer-term mortaltiy prognosis and the lessened likelihood for significant bias.

In this comprehensive benchmark prognostic study, we demonstrated that machine learning models provide more powerful one-year mortality prediction. Furthermore, we comprehensively assessed variable importance for each model in both the post-policy and pre-polciy eras, which may shed clinical insights for how we perform clinical decision-making for heart transplant surgeons. As clinical care practices continue to evolve, exploring computational time-to event algorithms to develop statistically powerful mortality risk prediction models is warranted.

BMI = body mass index; ECMO = extracorporeal membrane oxygenation; IABP = intra-aortic blood pump; ICU = intensive care unit; LOS = length of stay; OPTN = Organ Procurement and Transplantation Network; SRTR = Scientific Registry of Transplant Recipients; VAD = ventricular assist device

Authorship: Lathan Liou designed the study, performed the data analysis, and wrote the manuscript. EM, MAM, AB helped design the study and edit the manuscript. Laura Lehman, SS, SG, FJB, YW, and AC helped edit the manuscript.

Competing interests: The authors declare no conflicts of interest.

Acknowledgements: We would like to thank Annette Spooner for sharing her machine learning benchmark code with us and for helping with some of the technical components of the analysis.

Funding: Elizabeth Mostofsky is funded by Grant # K01AA027831 from the National Institute on Alcohol Abuse and Alcoholism (NIAAA). The content is solely the responsibility of the authors and does not necessarily represent the official views of the funding organizations.

Data Availability Statement: The data are available from SRTR upon request by emailing [email protected] (More information can be found at https://www.srtr.org/requesting-srtr-data/data-requests/). SRTR is a third-party organization that owns the data, which is why we are not allowed to share data per the Data Use Agreement that we signed with SRTR. The authors confirm they did not have any special privileges that other authors would not have.

Disclaimer: The data reported here have been supplied by the Hennepin Healthcare Research Institute (HHRI) as the contractor for the Scientific Registry of Transplant Recipients (SRTR). The interpretation and reporting of these data are the responsibility of the author(s) and in no way should be seen as an official policy of or interpretation by the SRTR or the U.S. Government.

Khush KK, Cherikh WS, Chambers DC, Harhay MO, Hayes D, Hsich E, et al. The International Thoracic Organ Transplant Registry of the International Society for Heart and Lung Transplantation: Thirty-sixth adult heart transplantation report — 2019; focus theme: Donor and recipient size match. J Heart Lung Transplant. 2019 Oct 1;38(10):1056–66.
Weiss ES, Allen JG, Kilic A, Russell SD, Baumgartner WA, Conte JV, et al. Development of a quantitative donor risk index to predict short-term mortality in orthotopic heart transplantation. J Heart Lung Transplant Off Publ Int Soc Heart Transplant. 2012 Mar;31(3):266–73.
Hsich EM, Blackstone EH, Thuita LW, McNamara DM, Rogers JG, Yancy CW, et al. Heart Transplantation: An In-Depth Survival Analysis. JACC Heart Fail. 2020 Jul;8(7):557–68.
Ayers B, Sandholm T, Gosev I, Prasad S, Kilic A. Using machine learning to improve survival prediction after heart transplantation. J Card Surg. 2021 Nov;36(11):4113–20.
Dani A, Heidel JS, Qiu T, Zhang Y, Ni Y, Hossain MM, et al. External validation and comparison of risk score models in pediatric heart transplants. Pediatr Transplant. 2021 Dec 8;e14204.
Kampaktsis PN, Tzani A, Doulamis IP, Moustakidis S, Drosou A, Diakos N, et al. State-of-the-art machine learning algorithms for the prediction of outcomes after contemporary heart transplantation: Results from the UNOS database. Clin Transplant. 2021 Aug;35(8):e14388.
George B, Seals S, Aban I. Survival analysis and regression models. J Nucl Cardiol Off Publ Am Soc Nucl Cardiol. 2014 Aug;21(4):686–94.
Adult heart allocation - OPTN [Internet]. [cited 2022 Jul 29]. Available from: https://optn.transplant.hrsa.gov/professionals/by-organ/heart-lung/adult-heart-allocation/
Lazenby KA, Narang N, Pelzer KM, Ran G, Parker WF. An updated estimate of posttransplant survival after implementation of the new donor heart allocation policy. Am J Transplant Off J Am Soc Transplant Am Soc Transpl Surg. 2022 Jun;22(6):1683–90.
Varshney AS, Berg DD, Katz JN, Baird-Zars VM, Bohula EA, Carnicelli AP, et al. Use of Temporary Mechanical Circulatory Support for Management of Cardiogenic Shock Before and After the United Network for Organ Sharing Donor Heart Allocation System Changes. JAMA Cardiol. 2020 Jun 1;5(6):703–8.
Zhang Z. Multiple imputation with multivariate imputation by chained equation (MICE) package. Ann Transl Med. 2016 Jan;4(2):30.
Kamalia MA, Ramamurthi A, Rein L, Mohammed A, Joyce DL. Detection of Seasonal Trends in National Donor Heart Availability Using the UNOS Dataset. J Card Fail. 2019 Aug 1;25(8, Supplement):S174.
Tibshirani R. The lasso method for variable selection in the Cox model. Stat Med. 1997 Feb 28;16(4):385–95.
Simon N, Friedman J, Hastie T, Tibshirani R. Regularization Paths for Cox’s Proportional Hazards Model via Coordinate Descent. J Stat Softw. 2011 Mar;39(5):1–13.
Friedman JH. Stochastic gradient boosting. Comput Stat Data Anal. 2002 Feb 28;38(4):367–78.
Chen T, Guestrin C. XGBoost: A Scalable Tree Boosting System. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining [Internet]. New York, NY, USA: Association for Computing Machinery; 2016 [cited 2022 Jul 29]. p. 785–94. (KDD ’16). Available from: https://doi.org/10.1145/2939672.2939785
Barnwal A, Cho H, Hocking T. Survival regression with accelerated failure time model in XGBoost. J Comput Graph Stat. 2022;31(4):1292–302.
Hemant Ishwaran, Udaya B. Kogalur, Eugene H. Blackstone, Michael S. Lauer. Random survival forests. Ann Appl Stat. 2008 Sep 1;2(3):841–60.
Bischl B, Lang M, Kotthoff L, Schiffner J, Richter J, Studerus E, et al. mlr: Machine Learning in R. J Mach Learn Res. 2016;17(170):1–5.
Strobl C, Boulesteix AL, Zeileis A, Hothorn T. Bias in random forest variable importance measures: Illustrations, sources and a solution. BMC Bioinformatics. 2007 Jan 25;8(1):25.
Weiss ES, Allen JG, Arnaoutakis GJ, George TJ, Russell SD, Shah AS, et al. Creation of a quantitative recipient risk index for mortality prediction after cardiac transplantation (IMPACT). Ann Thorac Surg. 2011 Sep;92(3):914–21; discussion 921-922.
Johnston LE, Grimm JC, Magruder JT, Shah AS. Development of a Transplantation Risk Index in Patients With Mechanical Circulatory Support: A Decision Support Tool. JACC Heart Fail. 2016 Apr;4(4):277–86.
Spooner A, Chen E, Sowmya A, Sachdev P, Kochan NA, Trollor J, et al. A comparison of machine learning methods for survival analysis of high-dimensional clinical data for dementia prediction. Sci Rep. 2020 Nov 23;10(1):20410.
Hsich EM, Thuita L, McNamara DM, Rogers JG, Valapour M, Goldberg LR, et al. Variables of importance in the Scientific Registry of Transplant Recipients database predictive of heart transplant waitlist mortality. Am J Transplant Off J Am Soc Transplant Am Soc Transpl Surg. 2019 Jul;19(7):2067–76.
Tian D, Yan HJ, Huang H, Zuo YJ, Liu MZ, Zhao J, et al. Machine Learning–Based Prognostic Model for Patients After Lung Transplantation. JAMA Netw Open. 2023 May 5;6(5):e2312022.
Mistiaen WP. Heart transplantation in patients with previous malignancy. An overview. Acta Cardiol. 2015 Apr;70(2):123–30.
Osorio‐Jaramillo E, Haasnoot GW, Kaider A, Schaefer A, Haberl T, Goekler J, et al. Molecular‐level HLA mismatch is associated with rejection and worsened graft survival in heart transplant recipients – a retrospective study. Transpl Int. 2020 Sep;33(9):1078–88.
Nam K, Jang EJ, Kim GH, Yhim HB, Lee H, Kim DH, et al. Perioperative red blood cell transfusion and mortality following heart transplantation: A retrospective nationwide population-based study between 2007 and 2016 in Korea. J Card Surg. 2019 Oct;34(10):927–32.
Khush KK, Ball RL. Great variability in donor heart acceptance practices across the United States. Am J Transplant Off J Am Soc Transplant Am Soc Transpl Surg. 2020 Jun;20(6):1582–96.
Hong KN, Iribarne A, Worku B, Takayama H, Gelijns AC, Naka Y, et al. Who is the high-risk recipient? Predicting mortality after heart transplant using pretransplant donor and recipient risk factors. Ann Thorac Surg. 2011 Aug;92(2):520–7; discussion 527.
Trivedi JR, Cheng A, Ising M, Lenneman A, Birks E, Slaughter MS. Heart Transplant Survival Based on Recipient and Donor Risk Scoring: A UNOS Database Analysis. ASAIO J Am Soc Artif Intern Organs 1992. 2016 Jun;62(3):297–301.
Trivedi JR, Slaughter MS. “Unintended” Consequences of Changes in Heart Transplant Allocation Policy: Impact on Practice Patterns. ASAIO J Am Soc Artif Intern Organs 1992. 2020 Feb;66(2):125–7.
Parker WF, Chung K, Anderson AS, Siegler M, Huang ES, Churpek MM. Practice Changes at U.S. Transplant Centers After the New Adult Heart Allocation Policy. J Am Coll Cardiol. 2020 Jun 16;75(23):2906–16.
Crawford TC, Magruder JT, Grimm JC, Kemp CD, Suarez-Pierre A, Zehr KJ, et al. The Paradoxical Relationship Between Donor Distance and Survival After Heart Transplantation. Ann Thorac Surg. 2017 May;103(5):1384–91.
Hoffman JRH, Larson EE, Rahaman Z, Absi T, Levack M, Balsara KR, et al. Impact of increased donor distances following adult heart allocation system changes: A single center review of 1-year outcomes. J Card Surg. 2021 Oct;36(10):3619–28.

No competing interests reported.

LiouSRTRpost2018mlSuppv3.docx

Download PDF

Version 1

posted

You are reading this latest preprint version

Survival Machine Learning Methods for Mortality Prediction After Heart Transplantation

Status:

Version 1

Abstract

Figures

INTRODUCTION

METHODS

RESULTS

DISCUSSION

CONCLUSIONS

Abbreviations

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1