Predicting Where Patients Will Be, Rather Than Just Seeing Where They Are
2020; Lippincott Williams & Wilkins; Volume: 141; Issue: 24 Linguagem: Inglês
10.1161/circulationaha.120.047571
ISSN1524-4539
AutoresYasbanoo Moayedi, Jeffrey J. Teuteberg,
Tópico(s)Artificial Intelligence in Healthcare
ResumoHomeCirculationVol. 141, No. 24Predicting Where Patients Will Be, Rather Than Just Seeing Where They Are Free AccessEditorialPDF/EPUBAboutView PDFView EPUBSections ToolsAdd to favoritesDownload citationsTrack citationsPermissions ShareShare onFacebookTwitterLinked InMendeleyReddit Jump toFree AccessEditorialPDF/EPUBPredicting Where Patients Will Be, Rather Than Just Seeing Where They AreEstablishing Trajectories of Cardiac Allograft Vasculopathy Yasbanoo Moayedi, MD and Jeffrey J. Teuteberg, MD Yasbanoo MoayediYasbanoo Moayedi Ted Rogers Centre of Excellence in Heart Function, Peter Munk Cardiac Centre, University Health Network, Toronto, Canada (Y.M.). and Jeffrey J. TeutebergJeffrey J. Teuteberg Jeffrey J. Teuteberg, MD, 300 Pasteur Dr, Stanford, CA 94305. Email E-mail Address: [email protected] https://orcid.org/0000-0002-8342-3222 Section of Heart Failure, Cardiac Transplant and Mechanical Circulatory Support, Stanford University, CA (J.J.T.). Originally published15 Jun 2020https://doi.org/10.1161/CIRCULATIONAHA.120.047571Circulation. 2020;141:1968–1970This article is a commentary on the followingIdentification and Characterization of Trajectories of Cardiac Allograft Vasculopathy After Heart TransplantationI skate to where the puck is going to be,not where it has been.—Wayne GretzkyArticle, see p 1954Despite significant strides in the field of heart transplantation, cardiac allograft vasculopathy (CAV) remains the leading long-term cause of death and retransplantation.1–3 CAV is a chronic fibroproliferative disease; its pathophysiology is related to the complex interplay of immunology and traditional coronary artery disease risk factors resulting in panarterial vascular changes.2 Unlike many posttransplant comorbidities, patients may remain asymptomatic despite severe progression of CAV. Given the diffuse and concentric nature of the vascular changes, conventional coronary angiography is of limited value in the early diagnosis of CAV, often underestimating the burden of disease. Additional testing in conjunction with angiography, such as intravascular ultrasound or optical coherence tomography, has improved early detection, but these testing options are not universally available and have not been equally applicable to children and adults.3,4 Furthermore, angiographic assessments are frequently limited by the presence of posttransplant renal insufficiency. Last, whereas statins improve posttransplant mortality, and proliferation signal inhibitors may slow the progression of CAV, programs may develop a diagnostic nihilism for ongoing invasive assessments for patients who are either intolerant or already on these therapies because of the lack of alternatives. When making decisions regarding CAV, we are often overly concerned with the last angiogram, stuck focusing on where the puck has been, rather than being able to incorporate a more comprehensive patient-centered assessment to understand the individual risk for CAV and better focus our efforts on where the puck is going to be.Machine learning (ML) has emerged as a potential means of assessing large, complex data sets by applying analytic algorithms that iteratively learn data to reveal novel patterns that may improve on traditional statistical methods of determining risk.5,6 Although transplant programs rely on protocols that are applied across an entire cohort of patients, the use of ML may allow physicians to better understand risk and to move to more individualized diagnostic and therapeutic strategies. In classic ML, data sets are divided into a training and testing cohort to build and subsequently evaluate the performance of a model. An unsupervised ML approach assesses patterns without predefined labels to bring order to the data set. However, a predictive model is only as good as the input data.5 Previous analyses using ML in the United Network of Organ Sharing registries for heart transplantation have not outperformed linear models.7 Although these databases are at an advantage with large numbers of patients, they often lack the detail and repeated measures found in relatively smaller institutional databases. In this issue of Circulation, Loupy et al8 have used 4 institutional databases to incorporate the aspects of ML in the assessment of CAV.The authors present a provocative study that attempts to phenotype CAV through an unsupervised ML approach by using the data of 2 European centers as a training set and the data of a single US center for validation. The study includes 1301 patients (815 European, 486 US) to identify 4 distinct CAV categories ranging from profile 1 (no CAV with no progression) to profile 4 (early development of CAV with significant progression). By using standard multinomial regression, they also identified 6 risk factors for these CAV phenotypes: donor age, donor sex, donor tobacco use, recipient low-density lipoprotein >1 g/L, class II donor-specific antibodies, and presence of moderate cellular rejection ≥2R at 1 year. As a result of this work, the authors provide an online risk stratification tool that may inform clinicians regarding the likelihood of a CAV risk profile. The strengths of the study include the robust and well-characterized clinical, histological, and immunologic nature of the data sets. Patients were monitored for a median of 7 years with very few missing data. Furthermore, these profiles have a face validity that largely corresponds with clinical practice. We commend the authors for their use of these novel techniques to tackle one of the most vexing long-term complications in cardiac transplantation.Using the referenced risk calculator, survival is listed for each profile in the Table. Survival curves start to diverge at 7 years, but there is little difference among profiles 1 through 3, despite the potential development of fairly severe CAV in profile 3 in comparison with profile 1. Therefore, although a patient may be categorized into profile 3 at 1 year, the implications for surveillance and therapy remain uncertain. Is the potential benefit of a conversion to a proliferation signal inhibitor worth the risk, given the inability to ascertain the likelihood of changing profiles over time? Even for patients who are assigned a low-risk category at the first annual visit, late-onset rejection and the later development of donor-specific antibodies may lead to rapid-onset CAV, which is not accounted for in the initial trajectory. To provide a more personalized prediction, it may be useful to allow input of serial data to assess the need for later monitoring and tailor surveillance to the patient's specific trajectory.Table. Survival Based on Cardiac Allograft Vasculopathy ProfileSurvival, %3-y5-y7-yProfile 110097.391.8Profile 210010096.1Profile 396.591.984.3Profile 495.686.572.1The study also relies on the International Society of Heart and Lung Transplantation consensus grading, which is insensitive to the development of early CAV. By definition, CAV1 is stenosis of <50% in the left main coronary artery or 70% in a primary vessel. Even at this stage there may already be diffuse vascular involvement that may subsequently limit the potential effectiveness of therapeutic interventions. The current model also does not appear to take into account other criteria of CAV such as left ventricular dysfunction or hemodynamic parameters that may modify the defined profile. In the future, prediction models could be further strengthened by incorporating baseline and 1-year assessment by intravascular ultrasound to account for donor-derived disease and the known impact of early changes in intimal thickness, which may be severe even in the presence of normal-appearing angiography.9 Furthermore, the current model categorizes the recipient based on the profile with the highest probability, which may be falsely reassuring. For example, a recipient with a history of elevated low-density lipoprotein, moderate cellular rejection at 1 year, and no donor-specific antibodies, who receives a 40-year-old male donor heart, has a 28.5% probability of profile 1, a 34% probability of profile 2, a 15% probability of profile 3, and a 22% probability of profile 4.10 Although deemed profile 2, there is a 66% probability that the recipient may not be that phenotype. Therefore, the model does not provide a personalized score, because it only describes which of the 4 profiles of CAV progression is most likely at a single time point. Once assigned to a CAV risk profile, there is no accounting for risks of death in the current mixed-methods model. Because of the inherent relation between CAV and posttransplant mortality, it is unclear how mortality would change the identification of distinct phenotypes.11Although these disease trajectories have limitations, the study provides a unique insight into CAV by providing the first description of potentially distinct patterns of disease progression. In the future, we hope that such collaborations will continue to provide the basis for the application of ML techniques to better understand complex disease processes such as CAV. Loupy et al should be applauded for bringing the heart transplant community 1 step closer to being able to skate to where the puck will be.AcknowledgmentsThe authors thank Dr Fan for the statistical review.DisclosuresDr Teuteberg serves on the speaking/advisory board for CareDx and Medtronic, consults for Abbott, is on the advisory board for Abiomed, and is a speaker for Paragonix. Dr Moayedi has nothing to disclose.FootnotesThe opinions expressed in this article are not necessarily those of the editors or of the American Heart Association.https://www.ahajournals.org/journal/circJeffrey J. Teuteberg, MD, 300 Pasteur Dr, Stanford, CA 94305. Email jeff.[email protected]eduReferences1. Khush KK, Cherikh WS, Chambers DC, Harhay MO, Hayes D, Hsich E, Meiser B, Potena L, Robinson A, Rossano JW, et al; International Society for Heart and Lung Transplantation. 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; 38:1056–1066. doi: 10.1016/j.healun.2019.08.004CrossrefMedlineGoogle Scholar2. Chih S, Chong AY, Mielniczuk LM, Bhatt DL, Beanlands RS. Allograft vasculopathy: the Achilles' heel of heart transplantation.J Am Coll Cardiol. 2016; 68:80–91. doi: 10.1016/j.jacc.2016.04.033CrossrefMedlineGoogle Scholar3. Mehra MR. The scourge and enigmatic journey of cardiac allograft vasculopathy.J Heart Lung Transplant. 2017; 36:1291–1293. doi: 10.1016/j.healun.2017.10.010CrossrefMedlineGoogle Scholar4. Clerkin KJ, Ali ZA, Mancini DM. New developments for the detection and treatment of cardiac vasculopathy.Curr Opin Cardiol. 2017; 32:316–325. doi: 10.1097/HCO.0000000000000388CrossrefMedlineGoogle Scholar5. Shameer K, Johnson KW, Glicksberg BS, Dudley JT, Sengupta PP. Machine learning in cardiovascular medicine: are we there yet?Heart. 2018; 104:1156–1164. doi: 10.1136/heartjnl-2017-311198CrossrefMedlineGoogle Scholar6. Rajkomar A, Dean J, Kohane I. Machine learning in medicine.N Engl J Med. 2019; 380:1347–1358. doi: 10.1056/NEJMra1814259CrossrefMedlineGoogle Scholar7. Miller PE, Pawar S, Vaccaro B, McCullough M, Rao P, Ghosh R, Warier P, Desai NR, Ahmad T. Predictive abilities of machine learning techniques may be limited by dataset characteristics: insights from the UNOS database.J Card Fail. 2019; 25:479–483. doi: 10.1016/j.cardfail.2019.01.018CrossrefMedlineGoogle Scholar8. Loupy A, Coutance G, Bonnet G, Van Keer J, Raynaud M, Aubert O, Bories M-C, Racapé M, Yoo D, Duong Van Huyen J-P, et al. Identification and characterization of trajectories of cardiac allograft vasculopathy after heart transplantation: a population-based Study.Circulation. 2020; 141:1954–1967. doi: 10.1161/CIRCULATIONAHA.119.044924LinkGoogle Scholar9. Kobashigawa JA, Tobis JM, Starling RC, Tuzcu EM, Smith AL, Valantine HA, Yeung AC, Mehra MR, Anzai H, Oeser BT, et al. Multicenter intravascular ultrasound validation study among heart transplant recipients: outcomes after five years.J Am Coll Cardiol. 2005; 45:1532–1537. doi: 10.1016/j.jacc.2005.02.035CrossrefMedlineGoogle Scholar10. Personalized Prediction of CAV Trajectory.https://transplant-prediction-system.shinyapps.io/CAV_trajectories/. Accessed April 30, 2020.Google Scholar11. Proust-Lima C, Philipps V, Liquet B. Estimation of extended mixed models using latent classes and latent processes: the R package lcmm.J Stat Software. 2017; 78. doi: 10.18637/jss.v078.i02CrossrefGoogle Scholar Previous Back to top Next FiguresReferencesRelatedDetailsCited ByPeyster E, Janowczyk A, Swamidoss A, Kethireddy S, Feldman M and Margulies K (2022) Computational Analysis of Routine Biopsies Improves Diagnosis and Prediction of Cardiac Allograft Vasculopathy, Circulation, 145:21, (1563-1577), Online publication date: 24-May-2022. Khoury M, Conway J, Gossett J, Edens E, Soto S, Cantor R, Koehl D, Barnes A, Exil V, Glass L, Kirklin J and Zuckerman W (2022) Cardiac allograft vasculopathy in pediatric heart transplant recipients does early-onset portend a worse prognosis?, The Journal of Heart and Lung Transplantation, 10.1016/j.healun.2022.01.012, 41:5, (578-588), Online publication date: 1-May-2022. Related articlesIdentification and Characterization of Trajectories of Cardiac Allograft Vasculopathy After Heart TransplantationAlexandre Loupy, et al. Circulation. 2020;141:1954-1967 June 16, 2020Vol 141, Issue 24 Advertisement Article InformationMetrics © 2020 American Heart Association, Inc.https://doi.org/10.1161/CIRCULATIONAHA.120.047571PMID: 32539613 Originally publishedJune 15, 2020 Keywordsunsupervised machine learningEditorialsrisk assessmentallograftsheart transplantationPDF download Advertisement
Referência(s)