Clinical decisions are mainly driven by the ability of physicians to apply risk stratification to patients. However, this task is difficult as it requires complex integration of numerous parameters and is impacted by patient heterogeneity. We sought to evaluate the ability of transplant physicians to predict the risk of long-term allograft failure and compare them to a validated artificial intelligence (AI) prediction algorithm.
We randomly selected 400 kidney transplant recipients from a qualified dataset of 4000 patients. For each patient, 44 features routinely collected during the first-year post-transplant were compiled in an electronic health record (EHR). We enrolled 9 transplant physicians at various career stages. At 1-year post-transplant, they blindly predicted the long-term graft survival with probabilities for each patient. Their predictions were compared with those of a validated prediction system (iBox). We assessed the determinants of each physician’s prediction using a random forest survival model.
Among the 400 patients included, 84 graft failures occurred at 7 years post-evaluation. The iBox system demonstrates the best predictive performance with a discrimination of 0.79 and a median calibration error of 5.79%, while physicians tend to overestimate the risk of graft failure. Physicians’ risk predictions show wide heterogeneity with a moderate intraclass correlation of 0.58. The determinants of physicians’ prediction are disparate, with poor agreement regardless of their clinical experience.
This study shows the overall limited performance and consistency of physicians to predict the risk of long-term graft failure, demonstrated by the superior performances of the iBox. This study supports the use of a companion tool to help physicians in their prognostic judgement and decision-making in clinical care.
The ability to predict the risk of a particular event is key to clinical decision-making, for example when predicting the risk of a poor outcome to help decide which patients should receive an organ transplant. Computer-based systems may help to improve risk prediction, particularly with the increasing volume and complexity of patient data available to clinicians. Here, we compare predictions of the risk of long-term kidney transplant failure made by clinicians with those made by our computer-based system (the iBox system). We observe that clinicians’ overall performance in predicting individual long-term outcomes is limited compared to the iBox system, and demonstrate wide variability in clinicians’ predictions, regardless of level of experience. Our findings support the use of the iBox system in the clinic to help clinicians predict outcomes and make decisions surrounding kidney transplants.
End-stage kidney disease (ESKD) represents a major global health burden with a prevalence exceeding 7 million1. Kidney transplantation is the best treatment for ESKD2 and is the most performed solid organ transplant in the world with more than 1,500,000 people living with a transplanted kidney according to the World Health Organization1,3. While short-term allograft survival has increased, improving long-term survival remains challenging4. The main limitations are the lack of new therapeutics as well as the absence of prediction systems that enable accurate risk stratification. Individual scoring systems that can risk stratify patients, serving as a companion tool for physicians, could improve patient management and facilitate the application of personalized medicine4.
However, accurately predicting patient outcomes can be challenging as physicians have to take into account numerous parameters from various sources, such as donor and recipient baseline characteristics5,6, follow-up parameters comprising immunological profile7,8,9, biomarkers, biopsy allograft phenotypes10, kidney function assessments11,12,13, treatment14, and parameters related to infections15, cancer16, and cardiovascular disease17. Consequently, accurately stratifying patient risk remains a difficult task for physicians9 and can lead to invasive examinations such as allograft biopsies or treatments with undesirable effects without any benefit for the patient.
For these reasons, several allograft failure prediction models have been developed18,19,20,21 to assist physicians in decision-making. The recently published iBox system is the most accurate and validated predictive algorithm in kidney transplantation to date and is currently undergoing review for regulatory endorsement by health authorities9. By adopting an integrative strategy using artificial intelligence to capture dependencies between numerous predictive factors, the iBox offers highly accurate allograft failure prediction performances and it has been validated in many centers, in distinct populations of transplant recipients, and in various clinical scenarios encountered in routine practice.
However, although existing risk prediction systems, including the iBox, have shown good predictive performances, none have been shown to outperform physicians. Before integrating these tools into clinical practice, it is necessary to investigate whether the predictions are more accurate than those made by physicians, which would provide a new perspective on the patient, potentially adjusting the prognostic judgment and therapeutic decisions.
Therefore, in this study, we aim to evaluate the ability of transplant physicians to predict and stratify the risk of long-term allograft failure and compare them to a validated artificial intelligence (AI) prediction algorithm. Transplant physicians were enrolled at distinct career stages, to assess their ability and consistency to predict and stratify long-term allograft failure, compared to the AI algorithm (the iBox system) based on the anonymized electronic health record of four-hundred kidney transplant recipients at one-year post-transplantation. Physicians’ predictions agreement and features that have led to their predictions were also investigated considering clinical experience. We show the overall limited performances of physicians to accurately predict individual long-term outcomes compared to the iBox system and demonstrate their wide heterogeneity of prognostic judgment, regardless of clinical experience. Taken together, we suggest the use of the iBox system to help physicians in the decision-making process.
From the 4000 kidney transplant recipients of a qualified prospective multicentric cohort, 400 patients with an evaluation available at one-year post-transplant were randomly selected as detailed in the flow chart (Fig. 1). This cohort includes consecutive patients over 18 years of age prospectively enrolled at the time of kidney transplantation from a living or deceased donor at Necker Hospital, Saint-Louis Hospital, Foch Hospital, and Toulouse Hospital between 1 January 2005, and 1 January 2014, in France and has been previously reported and used to develop the iBox prognostication system9.
All data from this cohort were anonymized and prospectively entered at the time of transplantation, at the time of post-transplant allograft biopsies, and at each transplant anniversary by using a standardized protocol. The electronic case report form (eCRF), includes features routinely collected in health care and kidney transplant, comprised demographic characteristics (including recipients’ comorbidities, age, sex, height, and weight), transplant characteristics (including Donor type, Donor comorbidities, immunological risk defined by circulating anti-HLA Donor specific antibodies at time of transplantation), biological features (including glomerular filtration rate estimated (eGFR) by the Modification of Diet in Renal Disease Study equation22, urine protein/creatinine ratio23, and circulating anti-HLA donor-specific antibodies (DSA) specificities and concentrations), and allograft histologic data (including elementary lesion scores and diagnoses interpreted according to the Banff international classification for allograft pathology24). Data were retrieved from the database on March 2018 and allograft outcomes were prospectively assessed in the Paris Transplant Group cohort up to 1 January 2021. All patients provided written informed consent at the time of transplantation. The institutional review board of the Paris Transplant Group approved the protocol of the study (NCT03474003, IRB: #000119258). This database has been approved by the National French Commission for Bioinformatics, Data, and Patient Liberty: CNIL registration number: 363505.
Patient risk evaluation, which was performed at one-year post-transplant according to the centers’ practices, comprised 44 transplant features included in the anonymized electronic health record (EHR). The features are detailed in Supplementary Table 1 and comprise: (i) Recipient characteristics including comorbidities, age, gender, number of years on dialysis, (ii) Donor characteristics including age, deceased/living, cause of death, history of hypertension or diabetes, (iii) Transplant characteristics including the number of HLA mismatches, cold ischemia time, induction therapy, delayed graft function, (iv) Biological features at time of evaluation including the kidney allograft function which was assessed by the glomerular filtration rate estimated (eGFR) by the Modification of Diet in Renal Disease Study equation22, and the proteinuria level using the urine protein/creatinine ratio23, (v) Immunological features at the time of transplant and evaluation, including circulating anti-HLA donor-specific antibodies against HLA-A, HLA-B, HLA-Cw, HLA-DR, HLA-DQ and HLA-DP which were assessed using single-antigen flow bead assays25 (mean fluorescence intensity and specificity), and (vi) Allograft pathology data according to the 2017 Banff international classification for allograft pathology24 (including elementary lesion scores and diagnoses).
We recruited nine independent physicians to review the EHR of the 400 anonymized patients to predict the risk of long-term allograft failure. Physicians were included in the study if they met several criteria including: (i) no involvement in the iBox development study, (ii) active involvement in daily kidney transplant patient care, (iii) acceptance of a training tutorial to use the online anonymized EHR, (iv) consent to use the EHR blinded from the outcome to be compared to an algorithm, (v) accept to review each of the 400 anonymized EHR. We then stratified the physicians according to their clinical experience to approximate the landscape of daily transplant care: three were residents in nephrology or transplant surgery, three were fellows in a general nephrology unit or a kidney transplant unit, and three were assistant professors or full professors (seniors) in a kidney transplant unit. Details about each physician and their clinical experience are summarized in Supplementary Table 2.
The outcome of interest was the individual prediction performances assessed by transplant physicians and the AI system respectively to predict the risk of long-term allograft failure. Kidney allograft failure was defined as a patient’s definitive return to dialysis or pre-emptive kidney retransplantation. Patients who died with a functioning allograft were censored at the time of death as patients with a functional allograft.
Individual allograft survival probabilities of the 400 included patients were computed using the iBox algorithm9, a validated AI system designed to predict the risk of long-term allograft failure up to seven years after evaluation, as previously described9. For each patient, the iBox was calculated based on the β regression coefficients from the iBox study using 8 features available in the EHR, including baseline characteristics (time from transplant to evaluation), functional features (eGFR and protein/creatinine ratio), immunological features (MFI of the immunodominant circulating anti-HLA donor-specific antibodies), and histological features including microcirculation inflammation (g and ptc Banff scores), interstitial inflammation and tubulitis (i and t Banff scores), transplant glomerulopathy (cg Banff score), and interstitial fibrosis/tubular atrophy (IFTA Banff score). These 8 clinically meaningful features were found to be independently associated with allograft failure among 44 features commonly and routinely collected in kidney transplant centers. The iBox algorithm has been externally validated in randomized clinical trials and multiple independent cohorts in Europe, North and South America9,26,27.
We created an online platform consisting of patient anonymized electronic health records (EHR) blinded to the allograft outcome and the iBox predictions. A visual of the online platform is shown in Supplementary Fig. 1.
The tasks of the physicians were as follows: (i) Read the EHR of each patient, (ii) Estimation of the long-term allograft survival at seven years after the time of evaluation, and selection of a percentage (0% = very high risk of allograft failure, 100% = very low risk of allograft failure), (iii) Selection of a score to further represent the risk of allograft failure from 0 (very high risk of allograft failure) to 10 (very low risk of allograft failure).
Continuous variables were described using means and standard deviations (SDs) or medians and the interquartile ranges. We compared means and proportions between groups using Student’s t-test, analysis of variance (Mann–Whitney test for MFI) or the chi-square test (or Fisher’s exact test if appropriate). Allograft survival was estimated using the Kaplan-Meier method. The duration of follow-up started with the patient risk evaluation (starting point) up to the date of kidney allograft failure, or the end of follow-up (1 January 2021). For patients who died with a functioning allograft, allograft survival was censored at the time of death as a functional allograft.
The accuracy of both the physicians’ predictions and the iBox prognostication system were assessed based on the discrimination using Harrell’s concordance index and with visual examination of calibration curves and median calibration error for each calibration plot (rms package in R). The median calibration error was assessed as the median of the absolute differences, for each quantile, between the predicted allograft failure, and the fraction of surviving allografts. Additionally, a decision curve analysis was performed to assess the clinical value using the net benefit of the iBox system compared with the physicians to predict allograft failure at seven years post-evaluation (dcurves package in R).
To identify the features driving each physician’s predictions, a random survival forest was performed for each physician. The mean decrease in accuracy was used to determine the relative importance of the first ten features that led to their predictions. We then ranked all features to obtain the overall top 10 features driving the physician’s judgment (randomForestSRC package in R). Fleiss kappa was used to measure inter-rater agreement between each physician’s ranking (irr package in R)28.
The inter-rater reliability was investigated using intraclass correlation (ICC) to evaluate the physician’s agreement to predict each individual patient percentage of allograft survival at seven years post-evaluation. A two-way random effect for multiple raters/measurements ICC model was fit using Stata29.
To assess the intra-rater reliability, the relationship between the two predictions made by each physician (probabilities and risk scores of allograft failure) were compared using a linear regression model for each physician. In addition, the inter-rater reliability to score each individual patient was also tested using a two-way random effect for multiple raters/measurements ICC model in Stata29.
All analyses were performed using R (version 3.6.3, R Foundation for Statistical Computing) and Stata (StataCorp. 2017. Stata Statistical Software: Release 15. College Station, TX: StataCorp LP). Values of p < 0.05 were considered significant, and all tests were two-tailed.
Further information on research design is available in the Nature Research Reporting Summary linked to this article.
The characteristics of the patients included at the time of transplantation are detailed in Supplementary Table 3 and were representative of the prospective multicentric cohort of 4000 kidney transplant recipients (see Supplementary Fig. 2). The mean patient age at the time of transplantation was 51.62 years (SD = 13.60), 224 were male (56.00%), and 341 allografts (85.25%) were from deceased donation. Regarding the immunological risk, 57 (14.25%) were previously transplanted, and 79 (19.75%) had circulating anti-HLA DSA at time of transplantation. At 7 years post-evaluation, 84 (21%) patients developed the primary outcome of allograft failure.
The AI system (iBox) and physicians’ predicted probabilities had distinct distributions (p < 0.001, Fig. 2). The medians [IQR] of the kidney allograft survival probabilities were 78% [61–89] for the iBox, 60% [40–80] for Resident #1, 70% [60–80] for Resident #2, 57% [41–77] for Resident #3, 60% [25–90] for Fellow #1, 40% [20–70] for Fellow #2, 77% [50–82] for Fellow #3, 69.5% [50–82] for Senior #1, 50% [30–80] for Senior #2 and 47% [25–72] for Senior #3 (Fig. 2). The kidney allograft survival prediction distribution of the subgroups of physicians with the same clinical experience were also significantly different (p < 0.001 for Residents, p < 0.001 for Fellows and p < 0.001 for Seniors physicians).
The discrimination at 7 years post-evaluation was higher for the iBox algorithm with a discrimination of 0.789 (Supplementary Fig. 3b), while the physicians had a lower and heterogenous discrimination performances with 0.638 for Resident #1, 0.754 for Resident #2, 0.755 for Resident #3, 0.771 for Fellow #1, 0.786 for Fellow #2, 0.736 for Fellow #3, 0.763 for Senior #1, 0.767 for Senior #2, 0.703 for Senior #3.
The calibration plots showed that the AI system predictions were more reliable. On average, physicians tended to overestimate the risk of graft failure at risk evaluation, regardless of the level of experience (Fig. 3). The percentage of median calibration errors (Supplementary Fig. 3a) was low for the iBox algorithm with 5.79% IQR [4.40–7.72] confirming an adequate calibration, while physicians had a higher percentage of, respectively, 18.10% IQR [7.33–29.68] for Resident #1, 9.60% IQR [7.05–10.66] for Resident #2, 19.16% IQR [15.65–24.36] for Resident #3, 19.44% IQR [5.35–29.24] for Fellow #1, 35.87% IQR [35.36–39.44] for Fellow #2, 8.62% IQR [2.92–9.31] for Fellow #3, 12.64% IQR [12.29–13.67] for Senior #1, 23.83% IQR[16.8–29.44] for Senior #2, 33.03% IQR [20.58–43.98] for Senior #3.
Decision curve analysis showed that the iBox has greater net benefit across a range of thresholds higher than 2% compared with most physicians (n = 8/9, 88.9%) and 10% for all physicians (Supplementary Fig. 4). At a threshold of 20% (percentage of allograft failure observed), the net benefit of the iBox system is that the model identified 46 more cases per 1000 without increasing the number of patients treated unnecessarily. Conversely, the net benefit was lower for all the physicians.
We used intraclass correlation (ICC) to compare a quantitative variable with multiple raters. The individual graft survival probabilities predicted by physicians showed wide heterogeneity with a moderate intraclass correlation of 0.58 95% CI [0.51–0.64] between all physicians. We also compared the intraclass correlation by considering the clinical experience and achievement of each physician. The inter-rater reliability remains poor with an ICC of 0.48 95% CI [0.39–0.56] for Residents and moderate for Fellows and Seniors physicians with an ICC of 0.61 95% CI [0.39–0.74] and 0.59 95% CI [0.45–0.69] respectively.
The most consistent feature between the physicians and the iBox system was eGFR with a Fleiss kappa of 0.75 (p < 0.001). The hierarchy of the features were broadly different across physicians with an overall Fleiss kappa of 0.13 (Supplementary Table 4). Several features constituting the iBox system, hence independently associated with the risk of long-term allograft failure, were not driving forces in the physicians’ prognostic judgment (Fig. 4). Indeed, apart from the eGFR, the other features present in the iBox score with a slight agreement between all physicians were the histological Banff scores of glomerulonephritis [g score] (Fleiss kappa 0.17) and peritubular capillaritis [ptc score] (Fleiss kappa 0.11). The circulating anti-HLA donor specific antibody status at time of the evaluation was not ranked by a third of the physicians (two residents and one senior, Fig. 4). We then stratified the agreement by their clinical experience. The agreement remained poor with a Fleiss kappa of 0.02 for residents, 0.20 for fellows and 0.10 for seniors (Supplementary Table 4). Among physicians with the same clinical experience, the features with the best agreement were expanded criteria donor (Fleiss kappa 0.46) and eGFR (Fleiss kappa 0.26) for residents, eGFR (Fleiss kappa 1.00) and donor age (Fleiss kappa 0.46) for fellows and eGFR (Fleiss kappa 1.00) only for seniors.
To assess physician reliability between estimating the percentage of allograft survival and scoring a risk of allograft failure at seven years post-evaluation, a linear regression was performed for each physician (Supplementary Fig. 5). A strong negative linear relationship corresponding to a high reliability between the two assessments was found for each physician with a R-squared of 0.905 for Resident #1 (p < 0.001), 0.766 for Resident #2 (p < 0.001), 0.860 for Resident #3 (p < 0.001), 0.985 for Fellow #1 (p < 0.001), 0.929 for Fellow #2 (p < 0.001), 0.973 for Fellow #3 (p < 0.001), 0.992 for Senior #1 (p < 0.001), 0.934 for Senior #2 (p < 0.0001) and 0.822 for Senior #3 (p < 0.001).
We also investigated the intraclass correlation of physicians to score the risk of allograft failure. The intraclass correlation (ICC) remained moderate at 0.55% CI [0.46–0.62] between all physicians, when taking into account the achievement of each physician, the results were consistent with an ICC of 0.52 95% CI [0.19–0.71] for Residents, 0.47 95% CI [0.41–0.53] for Fellows and 0.57 95% CI [0.42–0.68] for Seniors.
In this study, we investigated the prediction performances of an AI system and 9 transplant physicians with distinct clinical experiences, in assessing the risk of long-term allograft failure after kidney transplant. We showed that the iBox had better prediction performances than physicians, regardless of their experience. We also showed that physicians had limited performance, reproducibility, and consistency to predict the risk of long-term allograft failure.
Interestingly, few physicians had a discrimination close to the iBox but they all tended to overestimate the risk of allograft failure while the iBox showed a good discrimination and a strong, stable calibration. The predicted risk corresponds to the actual outcome for a large combination of predictor values. In addition, the predicted probabilities of long-term allograft failure were highly heterogeneous between physician estimates, while the iBox is stable.
We further supported this argument by ranking the feature importance for both the iBox and physicians. The physicians demonstrated a high heterogeneity in the choice of features that best predict the risk of long-term allograft failure. This result was not influenced by clinical experience, underscoring the possibility that this heterogeneity may be present in all physicians regardless of their experience. Overall, we demonstrated that physicians estimated that some key features independently associated with allograft failure described in the literature and included in the iBox were not, in their professional opinion, the most relevant driving forces6,30,31.
Therefore, as one given patient has one given risk of losing the allograft according to a spectrum of parameters, this disparity demonstrates that even if one physician may sometimes accurately predict the risk of a patient, other physicians are unlikely to have the same accuracy. This can lead to heterogeneity of practices for the same patient between physicians with potential invasive examinations or unnecessary treatments without benefit to the patient. Better predicting kidney allograft survival can help physicians improve risk stratification with reinforced surveillance for patients at high risk.
Overall, these findings illustrate that the iBox can inform physicians’ prognostic judgment and therefore decision-making and monitoring. As such, the iBox is a promising companion tool in daily transplant practice.
Kidney transplantation is a health care field representative of the quest for precision medicine over the past two decades32. Transplant physicians are overwhelmed with increasing data that are subject to many changes in definition and evaluation. For instance, the international Banff classification of allograft pathology has been updated every two years since 1991, making the interpretation of histological lesions increasingly complex33,34. Furthermore, to detect anti-HLA donor-specific antibodies, the Luminex single antigen bead assay technique is used worldwide, but remains difficult to interpret for physicians due to the use of different cut-offs and interpretations between laboratories25,35. Together, these ongoing changes represent a diverse knowledge that requires, for a physician, a long experience in transplant care and research to be correctly understood and integrated. Therefore in this context, the iBox, which suffers from less bias associated with memory and computation capability than humans, is likely to be of valuable assistance in transplant care.
More generally, this study reinforces the effort already made by researchers to compare machines to humans in the diagnosis or prognosis based on clinical data. This effort has often been focused on how machines could outperform physicians for image classification and disease diagnosis, but also more recently in patient prognostication of short-term outcomes36,37,38,39. However, to the best of our knowledge, this study is one of the first to compare long-term outcome predictions from a validated integrative prognostication system to physician predictions using EHR.
Despite its superior prediction performances, the iBox system will not replace physicians. The value of the iBox is its integration of a large spectrum of parameters from miscellaneous sources highly associated with the risk of long-term allograft failure in kidney transplantation9. However, it does not integrate the complexity of the physician-patient relationship, which involves many subtleties that contribute to decision-making. Further, additional specific data such as complications related to immunosuppressive treatment and events like cancer and infections have an important influence on clinical decisions. They are however not considered by the iBox system, although it may indirectly integrate the consequences of these events. Therefore, even though physicians predict with lower accuracy the risk of long-term allograft failure, they also have a large overview of the patient that cannot be currently reached by the machine. Thus, instead of opposing these two perspectives, the iBox should be considered as a companion tool that helps the physician in the evaluation of the patient, and thereby may serve as a support decision-making tool.
In conclusion, the present study demonstrates the overall limited performance of physicians to accurately predict individual risk of long-term allograft outcome compared with a validated AI prognostic system. This study also shows the potential added value of the iBox prognostication system to inform physicians in their prognostic judgment in kidney transplantation, supporting the use of computer assistance to help physicians in the decision-making process.
All source data to reproduce the main Figs. 2–4 and the Supplementary Figs. 3–5 are deposited into the synapse public repository40. Additional data to reproduce Supplementary Fig. 2 are available upon reasonable request. Technical appendix is available from the corresponding author at firstname.lastname@example.org. Study protocol is available on clinicaltrials.gov: NCT04918199.