Application of regularized regression to identify novel predictors of mortality in a cohort of hemodialysis patients

Werfel, Stanislas; Lorenz, Georg; Haller, Bernhard; Günthner, Roman; Matschkal, Julia; Braunisch, Matthias C.; Schaller, Carolin; Gundel, Peter; Kemmner, Stephan; Hayek, Salim S.; Nusshag, Christian; Reiser, Jochen; Moog, Philipp; Heemann, Uwe; Schmaderer, Christoph

doi:10.1038/s41598-021-88655-0

Application of regularized regression to identify novel predictors of mortality in a cohort of hemodialysis patients

Article
Open access
Published: 29 April 2021

Volume 11, article number 9287, (2021)
Cite this article

Download PDF

You have full access to this open access article

Scientific Reports

Application of regularized regression to identify novel predictors of mortality in a cohort of hemodialysis patients

Download PDF

Stanislas Werfel¹^na1,
Georg Lorenz¹^na1,
Bernhard Haller²,
Roman Günthner¹,
Julia Matschkal¹,
Matthias C. Braunisch¹,
Carolin Schaller¹,
Peter Gundel¹,
Stephan Kemmner^1,3,
Salim S. Hayek⁴,
Christian Nusshag^5,6,
Jochen Reiser⁵,
Philipp Moog¹,
Uwe Heemann¹ &
…
Christoph Schmaderer¹

917 Accesses
Explore all metrics

Abstract

Cohort studies often provide a large array of data on study participants. The techniques of statistical learning can allow an efficient way to analyze large datasets in order to uncover previously unknown, clinically relevant predictors of morbidity or mortality. We applied a combination of elastic net penalized Cox regression and stability selection with the aim of identifying novel predictors of mortality in a cohort of prevalent hemodialysis patients. In our analysis we included 475 patients from the “rISk strAtification in end-stage Renal disease” (ISAR) study, who we split into derivation and confirmation cohorts. A wide array of examinations was available for study participants, resulting in over a hundred potential predictors. In the selection approach many of the well established predictors were retrieved in the derivation cohort. Additionally, the serum levels of IL-12p70 and AST were selected as mortality predictors and confirmed in the withheld subgroup. High IL-12p70 levels were specifically prognostic of infection-related mortality. In summary, we demonstrate an approach how statistical learning can be applied to a cohort study to derive novel hypotheses in a data-driven way. Our results suggest a novel role of IL-12p70 in infection-related mortality, while AST is a promising additional biomarker in patients undergoing hemodialysis.

Dynamic personalized risk prediction in chronic heart failure patients: a longitudinal, clinical investigation of 92 biomarkers (Bio-SHiFT study)

Article Open access 18 February 2022

Nomogram for Predicting Cardiovascular Mortality in Incident Peritoneal Dialysis Patients: An Observational Study

Article Open access 24 October 2017

Development and validation of a model for predicting the risk of cardiovascular events in maintenance hemodialysis patients

Article Open access 21 March 2024

Introduction

End-stage renal disease (ESRD) with hemodialysis treatment continues to be associated with dramatically increased morbidity and mortality. While cardiovascular mortality represents the prevalent cause of death in these patients, death due to other causes such as infection is also more frequent than in the general population¹. Several risk factors of cardiovascular and all-cause mortality have been identified and risk scores have been successfully established based on these^2,3,4,5,6,7. Beyond the quantification of a personal prognosis, the identification of novel risk factors can lead to hypothesis generation for further research, novel therapeutic interventions and finally lead to improvements in patient outcomes.

Machine learning approaches for survival analysis can allow efficient selection of the most relevant risk factors even when a large number of potential predictors for only a relatively small patient population is available^{Methods” (“All-cause model”, black curve). Solid lines represent the addition of AST (ln-transformed, fitted by a spline) and/or IL-12p70 (high group) to the all-cause model. Dashed lines represent the all-cause model after removal of the indicated known predictors. Dots in (b) represent AUC values for individual imputed datasets for each model (see “Methods”). Lines in (a) and (b) represent smoothed conditional means. Bootstrap** was performed on the total cohort with non-missing values for (a) and on the total cohort (n = 475) with missing values imputes as described in “Methods” section for (b). Tr., transformation; spl., spline fit; Catheter, use of catheter for dialysis.}

Full size image

The AUCs for these single predictors were comparable to those of other established predictors such as serum albumin and the use of a catheter for dialysis, though not as high as some of the top predictors for mortality, age and adapted comorbidity score (Fig. 5a).

Analysis of multivariable models showed an additive effect of the new predictors on our previously established model⁵ (Fig. 5b, see “Methods” for included predictors). In fact, inclusion of both AST and IL-12p70 (green line in Fig. 5b) was superior to albumin or catheter use and had a similar additive value on AUC as the strong predictors age and adapted CCI within the first year of follow up (orange and blue dashed lines in Fig. 4b left panel).

Discussion

Here, we applied techniques of statistical learning to analyze a cohort study of dialysis patients. Although similar techniques have been previously applied in this population in order to improve mortality prediction^18,19,20, our approach differs in that our main goal was not to derive an overall prediction model or to address a specific hypothesis, but rather to find novel promising candidates as predictors in a setting with a large amount of available characteristics for a relatively small study population. Many of the well-known and established prognostic factors for mortality were recovered, while a relatively modest number of 317 patients (with 112 events during follow-up) were used for selection among ≈ 110 predictors. We were then able to confirm significant associations in the withheld smaller confirmation group for most of the factors obtained after stability selection using univariate Cox regression. These findings underline the overall validity of the selection process in order to identify important mortality predictors. A notable observation was thereby the identification of a previously unknown association of IL-12p70 with infection-associated mortality.

IL-12p70 is the active heterodimer form of IL-12, which consists of the IL-12p35 and IL-12p40 subunits. While IL-12p35 is thought to be specific for IL-12, IL-12p40 also associates with IL-23p19 to form active IL-23. IL-12 activity has been implicated in the immunologic response to pathogens and tumor cells, as well as in autoimmune disease^21,22. Levels of IL-12p70 are known to be relatively low in the serum of healthy individuals²³ and were below the detection limit for our assay in the majority of the study population. Contradictory reports concerning IL-12 activity exist for dialysis patients. One study, published shortly after the discovery of this cytokine, observed an increased level of IL-12 in dialysis patients compared to healthy controls²⁴, a finding that could possibly be explained by increased IL-12 release from peripheral blood mononuclear cells (PBMCs) as an inflammatory response to the dialysis membrane²⁵. This study also reported, contrary to our observation, that increased IL-12 levels were associated with improved survival in dialysis patients²⁴. However, in a more recent report assessing specifically the levels of IL-12p40 and IL-12p70, a significant increase of IL-12p40 was observed, while no increase was reported for IL-12p70 in dialysis patients vs. healthy controls²⁶. The discrepancy between our finding and those of Kimmel et al.²⁴ could therefore be explained by the specificity for IL-12p70 in our assay. Alternatively, since we categorized data and compared individuals with 20% highest levels to others, the effects of IL-12 might be non-linear, in that they might be protective up to a certain level and detrimental or indicative of an additional underlying disease (e.g. autoimmune) above that level. Further studies employing more sensitive IL-12p70 and IL-12p40 measurements would be of interest to solve this question. Having high levels of IL-12p70 was one of the most stable predictors for increased infection-associated mortality in our derivation cohort. Although a causal relationship cannot be established by the applied approach, this close correlation suggests that it may be a causal factor or that it may be closely linked to a process causing mortality in hemodialysis patients. Further studies into the underlying pathophysiology would be highly interesting to elucidate this relationship and clarify if modulation of IL-12 activity may serve as a therapeutic intervention.

Another stable predictor which was obtained in our analysis is AST and we further looked into this association because until now it has received little attention in dialysis patients. AST is often used in clinical practice to assess liver damage, however unlike ALT it is less specific and is also elevated in other conditions, such as myocardial infarction or skeletal muscle damage^27,28. Levels of AST have been shown to be predictive of both all-cause as well as specifically cardiovascular mortality (potentially due to an association with the metabolic syndrome) in the general population^29,30,31,32. While on average a decreased AST activity in dialysis patients compared to the general population was observed long ago^33,34, we could only find one previous study analyzing its association with mortality in dialysis patients¹⁷ (in this case with all-cause mortality). Our results confirm these findings and suggest a potential association with cardiovascular disease in dialysis patients. An association with liver disease may explain an increased mortality due to "other" causes in patients with high AST values, similar as was observed for increased ALT values (Figure S2). The finding that ALT levels do not show a similar association with cardiovascular mortality, however, suggests that there may be an alternative relevant pathomechanism or a source of AST elevation outside the liver (such as the heart) which is worthy of further investigation. Although our study is not sufficiently powered to further elucidate the pathophysiology behind AST variation (e.g. only 5% of participants had a known liver disease based on medical records) our results underline the prognostic value of AST and support the previously expressed need for different reference ranges in hemodialysis patients³⁵.

When assessing the additional value of AST and IL-12p70 for the prediction of mortality in dialysis patients using time-dependent ROC analysis, we found an additive value of these predictors. Although in the multivariable model the numeric gain was relatively modest, it has to be compared to other established variables, because the model already incorporated many important predictors. We observed a similar effect of AST and IL-12p70 as e.g. for known independent predictors albumin and catheter use for dialysis. The added effect of these two predictors together was similar in the first year to some of the strongest predictors: age and adapted CCI score.

One of the limitations of the applied penalized Cox regression is that it assumes linear associations with the outcome (as does regular Cox regression). This also means that single high-leverage points, which can represent false measurements or extraordinary conditions, may have a strong effect on selection (or omission) of variables. Here we addressed these issues by “dummy coding” quantitative measurements to nominal variables which represent top or lowest quintiles for specific values. Although such thresholding may influence the selection due to loss of information, we in fact observed that most known ordinal/continuous predictors for mortality in dialysis patients passed our selection criteria (i.e. hsCRP, hemoglobin, creatinine, triglycerides, IL-6, YKL-40, AST). However, one notable exception is serum albumin, which was not retained despite a well-documented correlation in the literature. Decreased albumin levels are both observed as part of malnutrition as well as inflammation (anti-acute phase protein), also summarized as dialysis-associated inflammation-malnutrition complex³⁶. There is also a known negative correlation between IL-6 and albumin^37,38. Other retained predictors are also associated with inflammation-malnutrition in ESRD patients, such as hsCRP, creatinine or lipids³⁶. It is therefore likely that, next to the thresholding approach, such predictors have additionally diminished the explanatory value of albumin for mortality prediction, as has been reported previously³⁹. A similar collinearity of hsCRP with IL-6^40,41 may have reduced the predictive value of the former in the case of infection-associated mortality, yet it was selected as a predictor for all-cause mortality where more events were available and potentially due to its described additional role as a predictor of malnutrition⁴².

We strongly assume that including more patients in the derivation cohort would further improve factor selection, while in our study this dataset was relatively small (i.e. 317 patients). Having a larger cohort would also allow to include variable interactions in order to identify subgroup-specific effects, something that was not accounted for in our approach. Further limitations of our study are missing values for some of the variables and unknown cause of death for some of the patients. We expect that a relevant portion of missing values, as these were set to the reference level (see “Methods”), will reduce the selection stability of a variable and therefore typically lead to a more conservative selection. An additional conceptual limitation is that only the variables at baseline could be used in the current analysis. Yet, the evolution of biomarkers over time in dialysis patients and the course of chronic kidney disease may provide additional important insights and improve predictive performance.

In summary, we successfully applied techniques of statistical learning as an exploratory approach to high-dimensional data from a cohort of dialysis patients. Despite the relatively modest number of patients, we thereby could retrieve most of the established risk markers and additionally identify promising candidates for novel predictors. The described techniques can be considered complementary, yet not as competitors to widely-applied and state of the art hypothesis-driven approaches. Similar techniques, as they become more accessible, may gain importance in routine exploratory study analysis in order to gain additional insights into the underlying risk-determining factors.

Methods

Subjects/study population and data collection

The clinical trial was registered under ClinicalTrials.gov identifier: NCT01152892. For detailed information please refer to the study protocol¹². In brief, we included stable HD patients who were at least 18 years of age, had an HD vintage of at least 90 days, and provided written and informed consent. Malignant disease with a life expectancy shorter than 2 years had been the only relevant exclusion criterion. The present study included 475 patients, for whom sufficient unthawed sera were available for cytokine quantification. The study was approved by the ethics committees of the Klinikum rechts der Isar, Technical University Munich, and by the Bavarian State Board of Physicians. It was performed in accordance with the Declaration of Helsinki.

Age, patient characteristics, dialysis modalities, comorbidities and medication were assessed from medical records. Comorbidities were recorded at inclusion and updated during the 2 years follow up as previously described^5,43. An adapted CCI was calculated according to Liu et. al.¹³ (version without ESRD cause).

Cause of death was attributed to the following domains: cardiovascular, infection associated and other (or unknown) of which the cardiovascular domain had been prespecified in the study protocol¹². For further details also see Lorenz et al.⁵ Medical records and interviews of attending physicians, and relatives were used by the ISAR study physician board to ascertain death and assigned each case an underlying cause of death. When no agreement on cause of death could be reached the cause of death was defined as unknown (n = 28).

Serumbiochemistry and functional studies

Serum was collected prior to a midweek dialysis session at inclusion and after the exclusion of active infection. Patients that had undergone surgery within the past 2 weeks were included at a later timepoint. Serum was centrifuged after 30 min of resting at room temperature, aliquoted and stored at – 80 °C. Routine laboratory analyses were assessed by International Organization for Standardization–certified laboratories. YKL-40 levels were measured in duplicates using a commercial enzyme-linked immunosorbent assay (ELSIA) kit (R&D Systems, Inc., Minneapolis, MN). The remaining inflammatory mediators were assessed using BD Flex-sets (Becton Dickinson, Heidelberg, Germany) on a FACS Canto II (Becton Dickinson) and BD Diva software (Becton Dickinson) according to manufacturer’s instructions. Theoretical limits of detection were indicated by the manufacturer between 0.6 and 1.6 pg/mL.

Ambulatory 24-h blood pressure (BP) and pulse wave velocity (PWV) measurements were carried out as previously described⁴⁴ using a Mobil-O-Graph 24 h-PWA (I.E.M. GmbH, Stolberg, Germany). Retinal vessel analysis was performed using Static Vessel Analyzer (Imedos Systems UG, Jena, Germany). Fundus images of one eye were gathered and classified into four categories according to quality. Two images with the highest qualitative category were analyzed and means of central retinal arteriolar (CRAE) and central retinal venular equivalent (CRVE) as well as arteriolar-to-venular diameter ratio (AVR) were calculated as previously described⁴⁵.

Statistical analyses

Statistical analysis was performed in R version 3.6.3⁴⁶.

In order to be able to test the selected predictors on patients whose data was not used for variable selection, we split the cohort into a derivation and confirmation datasets. To this end we performed 1000 random splits in the ratio of 2/3 to 1/3 and calculated for each split the p-value for difference in the matching variables (age, adapted CCI, gender, IL-6 and mortality as indicated in Table 1) using Chi-squared test for nominal and Mann–Whitney test for ordinal or continuous variables. We then selected the split where the two groups were most likely similar (based on where the lowest p-value of the tests was the highest).

For stable variable selection we applied elastic net regularized regression on the derivation subcohort on all predictors listed in Table S1 simultaneously. In order to address possible non-linearity, ordinal and continuous variables (with the exception of age for which a linear relationship can be expected and which was expressed in decades) were dummy coded as “high” and “low” group by assigning for each patient whether he falls into the top or bottom quintile for the specific variable. Ties were assigned randomly. Values not falling into top or bottom quintile and missing values were assigned 0 in both dummy variables. Missing values in nominal variables were also assigned to the reference category (meaning absence of a potential risk factor).

In a first step we performed elastic net¹⁰ Cox regression separately for all-cause mortality, cardiovascular and infection-associated mortality with 20-fold cross-validation using 15 equally spaced alpha values in the range of 0 to 1 (here alpha = 0 equals ridge regression and alpha = 1 equals lasso regression). We selected for each alpha the lambda with the lowest cross-validation error (Figure S1a). We then selected for each regression the non-zero alpha value with the lowest cross validation error (vertical dashed lines in Figure S1a). Alpha of zero would not result in any variable selection (as it equals ridge regression) and was therefore not chosen. Next, we used the chosen alpha values in a stability selection approach as described by Meinshausen et Bühlmann¹¹ and implemented in R by Sill et al.⁴⁷ For this preselection approach as a means of hypothesis generation we used 100 subsamples of the derivation cohort, a “weakness” of 0.8 and a (maximal) per-comparison error rate of 10%. All variables passing the selection criteria are reported in Table 2.

For spline fitting we used penalized splines^48,49 with four degrees of freedom. For the analysis in Fig. 4 and Figure S2 one outlier with AST of > 600 U/l and ALT > 900 U/l (suggestive of acute hepatopathy) was removed.

The multivariable Cox model for all-cause mortality (Fig. 4b, Figure S2 and Fig. 5b) contained the following predictors (as described previously⁵): age, gender, BMI, adapted CCI, history of MI, rheumatic disease, laboratory parameters (phosphate, calcium, albumin), Kt/V, HD (not hemodiafiltration), usage of central venous catheter for dialysis, oral anticoagulation, IL-6 (ln-transformed) and YKL-40 (ln-transformed). The multivariable model for cardiovascular mortality contained: age, gender, BMI, adapted CCI, history of MI, rheumatic disease, usage of central venous catheter for dialysis, ultrafiltration per kilogram of dry weight, IL-6 (ln-transformed) and YKL-40 (ln-transformed). The model for infection-associated mortality contained age, gender, BMI, adapted CCI, history of MI, Albumin, usage of central venous catheter for dialysis, hsCRP (ln-transformed) and IL-6 (ln-transformed). For other mortality causes we used a minimal model which included: age, gender, BMI, adapted CCI and IL-6 (ln-transformed).

Time-dependent ROC analysis for Cox regression was performed by first fitting uni- (Fig. 5a) or multivariable (Fig. 5b) models to a bootstrap sample of the total study population and then calculating incident/dynamic AUCs⁵⁰ on the out of bootstrap sample. The bootstrap procedure was performed 1000 times and the resulting AUCs were averaged per month. For multivariable analysis missing values were randomly sampled from the available values for the respective variable and a total of 25 thus imputed datasets were generated. AUCs for each of these datasets were calculated as described above (these are presented as dots in Fig. 5b).

References

de Jager, D. J. et al. Cardiovascular and noncardiovascular mortality among patients starting dialysis. JAMA 302, 1782–1789 (2009).
Article PubMed Google Scholar
Anker, S. D. et al. Development and validation of cardiovascular risk scores for haemodialysis patients. Int. J. Cardiol. 216, 68–77 (2016).
Article PubMed Google Scholar
Thamer, M. et al. Predicting early death among elderly dialysis patients: Development and validation of a risk score to assist shared decision making for dialysis initiation. Am. J. Kidney Dis. 66, 1024–1032 (2015).
Article PubMed PubMed Central Google Scholar
Ma, L. & Zhao, S. Risk factors for mortality in patients undergoing hemodialysis: A systematic review and meta-analysis. Int. J. Cardiol. 238, 151–158 (2017).
Article PubMed Google Scholar
Lorenz, G. et al. Mortality prediction in stable hemodialysis patients is re fi ned by YKL-40, a 40-kDa glycoprotein. Kidney Int. 93, 221–230 (2018).
Article CAS PubMed Google Scholar
Günthner, R. et al. Impaired retinal vessel dilation predicts mortality in end-stage renal disease. Circ. Res. 124, 1796–1807 (2019).
Article CAS Google Scholar
Floege, J. et al. Development and validation of a predictive mortality risk score from a European hemodialysis cohort. Kidney Int. 87, 996–1008 (2015).
Article CAS PubMed PubMed Central Google Scholar
Wang, P., Li, Y. & Reddy, C. K. Machine Learning for Survival Analysis: A Survey. ar**v:1708.04649[cs.LG] (2017) [cs.LG].
Deo, R. C. Machine learning in medicine. Circulation 132, 1920–1930 (2015).
Article PubMed PubMed Central Google Scholar
Zou, H. & Hastie, T. Regularization and variable selection via the elastic net. J. R. Stat. Soc. Ser. B. Stat. Methodol. 67, 301–320 (2005).
Article MathSciNet MATH Google Scholar
Meinshausen, N. & Bühlmann, P. Stability selection. J. R. Stat. Soc. Ser. B Stat. Methodol. 72, 417–473 (2010).
Article MathSciNet MATH Google Scholar
Schmaderer, C. et al. Rationale and study design of the prospective, longitudinal, observational cohort study “rISk strAtification in end-stage renal disease” (ISAR) study. BMC Nephrol. 17, 1–12 (2016).
Article Google Scholar
Liu, J., Huang, Z., Gilbertson, D. T., Foley, R. N. & Collins, A. J. An improved comorbidity index for outcome analyses among dialysis patients. Kidney Int. 77, 141–151 (2010).
Article PubMed Google Scholar
Cohen, L. M., Ruthazer, R., Moss, A. H. & Germain, M. J. Predicting six-month mortality for patients who are on maintenance hemodialysis. Clin. J. Am. Soc. Nephrol. 5, 72–79 (2010).
Article PubMed PubMed Central Google Scholar
Keskar, V. & Sood, M. M. Use of Oral Anticoagulation in the Management of Atrial Fibrillation in Patients with ESRD : Con. 2085–2092 (2016) https://doi.org/10.2215/CJN.03200316.
Kilpatrick, R. D. et al. Association between Serum Lipids and Survival in Hemodialysis Patients and Impact of Race. 293–303 (2004) https://doi.org/10.1681/ASN.2006070795.
Ravel, V. et al. Association of aspartate aminotransferase with mortality in hemodialysis patients. Nephrol. Dial. Transplant 31, 814–822 (2016).
Article CAS PubMed Google Scholar
Goldstein, B. A., Pencina, M. J., Montez-Rath, M. E. & Winkelmayer, W. C. Predicting mortality over different time horizons: Which data elements are needed?. J. Am. Med. Informatics Assoc. 24, 176–181 (2017).
Article Google Scholar
Friedman, A. N. et al. Fatty acids and other risk factors for sudden cardiac death in patients starting hemodialysis. Am. J. Nephrol. 38, 12–18 (2013).
Article CAS PubMed Google Scholar
Goldstein, B. A., Chang, T. I., Mitani, A. A., Assimes, T. L. & Winkelmayer, W. C. Near-term prediction of sudden cardiac death in older hemodialysis patients using electronic health records. Clin. J. Am. Soc. Nephrol. 9, 82–91 (2014).
Article PubMed Google Scholar
Tait Wojno, E. D., Hunter, C. A. & Stumhofer, J. S. The immunobiology of the interleukin-12 family: Room for discovery. Immunity 50, 851–870 (2019).
Article CAS PubMed Google Scholar
Liu, J. et al. Interleukin-12: An update on its immunological activities, signaling and regulation of gene expression. Curr. Immunol. Rev. 1, 119–137 (2005).
Article CAS PubMed PubMed Central Google Scholar
Rea, I. M., McNerlan, S. E. & Alexander, H. D. Total serum IL-12 and IL-12p40, but not IL-12p70, are increased in the serum of older subjects; Relationship to CD3+ and NK subsets. Cytokine 12, 156–159 (2000).
Article CAS PubMed Google Scholar
Kimmel, P. L. et al. Immunologic function and survival in hemodialysis patients. Kidney Int. 54, 236–244 (1998).
Article CAS PubMed PubMed Central Google Scholar
Memoli, B., Marzano, L., Bisesti, V., Andreucci, M. & Guida, B. Hemodialysis-related lymphomononuclear release of interleukin-12 in patients with end-stage renal disease. J. Am. Soc. Nephrol. 10, 2171–2176 (1999).
Article CAS PubMed Google Scholar
Mansouri, L., Paulsson, J. M., Moshfegh, A., Jacobson, S. H. & Lundahl, J. Leukocyte proliferation and immune modulator production in patients with chronic kidney disease. PLoS ONE 8, e73141 (2013).
Article ADS CAS PubMed PubMed Central Google Scholar
Kim, W. R., Flamm, S. L., Di Bisceglie, A. M., Bodenheimer, H. C. & Public Policy Committee of the American Association for the Study of Liver Disease. Serum activity of alanine aminotransferase (ALT) as an indicator of health and disease. Hepatology 47, 1363–1370 (2008).
Article CAS PubMed Google Scholar
Rej, R. Aminotransferases in disease. Clin. Lab. Med. 9, 667–687 (1989).
Article CAS PubMed Google Scholar
Hyeon, C. K. et al. Normal serum aminotransferase concentration and risk of mortality from liver diseases: Prospective cohort study. BMJ 328, 983 (2004).
Article Google Scholar
Lee, T. H., Kim, W. R., Benson, J. T., Therneau, T. M. & Melton, L. J. Serum aminotransferase activity and mortality risk in a United States community. Hepatology 47, 880–887 (2008).
Article PubMed Google Scholar
Lee, H. et al. Association between change in serum aminotransferase and mortality. Medicine 95, e3158 (2016).
Article CAS PubMed PubMed Central Google Scholar
Oren, R. Serum liver enzymes—Should we count on them?. Liver Int. 34, 171–173 (2014).
Article CAS PubMed Google Scholar
Wolf, P. L., Williams, D., Coplon, N. & Coulson, A. S. Low aspartate transaminase activity in serum of patients undergoing chronic hemodialysis. Clin. Chem. 18, 567–568 (1972).
Article CAS PubMed Google Scholar
Sette, L. H. B. C. & de Almeida Lopes, E. P. Liver enzymes serum levels in patients with chronic kidney disease on hemodialysis: A comprehensive review. Clinics 69, 271–278 (2014).
Article PubMed PubMed Central Google Scholar
Ray, L., Nanda, S. K., Chatterjee, A., Sarangi, R. & Ganguly, S. A comparative study of serum aminotransferases in chronic kidney disease with and without end-stage renal disease: Need for new reference ranges. Int. J. Appl. Basic Med. Res. 5, 31–35 (2015).
Article CAS PubMed PubMed Central Google Scholar
Kalantar-Zadeh, K., Ikizler, T. A., Block, G., Avram, M. M. & Kopple, J. D. Malnutrition-inflammation complex syndrome in dialysis patients: Causes and consequences. Am. J. Kidney Dis. 42, 864–881 (2003).
Article PubMed Google Scholar
Spoto, B. et al. Association of IL-6 and a functional polymorphism in the IL-6 gene with cardiovascular events in patients with CKD. Clin. J. Am. Soc. Nephrol. 10, 232–240 (2015).
Article CAS PubMed Google Scholar
Bologa, R. M. et al. Interleukin-6 predicts hypoalbuminemia, hypocholesterolemia, and mortality in hemodialysis patients. Am. J. Kidney Dis. 32, 107–114 (1998).
Article CAS PubMed Google Scholar
Sun, J. et al. Biomarkers of cardiovascular disease and mortality risk in patients with advanced CKD. Clin. J. Am. Soc. Nephrol. 11, 1163–1172 (2016).
Article CAS PubMed PubMed Central Google Scholar
Panichi, V. et al. Interleukin-6 is a stronger predictor of total and cardiovascular mortality than C-reactive protein in haemodialysis patients. Nephrol. Dial. Transplant. 19, 1154–1160 (2004).
Article CAS PubMed Google Scholar
Wetmore, J. B. et al. Associations of interleukin-6, C-reactive protein and serum amyloid A with mortality in haemodialysis patients. Nephrology 13, 593–600 (2008).
Article CAS PubMed Google Scholar
Honda, H. et al. Serum albumin, C-reactive protein, interleukin 6, and fetuin a as predictors of malnutrition, cardiovascular disease, and mortality in patients with ESRD. Am. J. Kidney Dis. 47, 139–148 (2006).
Article CAS PubMed Google Scholar
Lorenz, G. et al. Worsening calcification propensity precedes all-cause and cardiovascular mortality in haemodialyzed patients. Sci. Rep. 7, 1–11 (2017).
Article CAS Google Scholar
Matschkal, J. et al. Comparison of 24-hour and office pulse wave velocity for prediction of mortality in hemodialysis patients. Am. J. Nephrol. 49, 317–327 (2019).
Article PubMed Google Scholar
Seidelmann, S. B. et al. Retinal vessel calibers in predicting long-term cardiovascular outcomes. Circulation 134, 1328–1338 (2016).
Article PubMed PubMed Central Google Scholar
R Core Team. R: A Language and Environment for Statistical Computing. (R Foundation for Statistical Computing, 2020).
Sill, M., Hielscher, T., Becker, N. & Zucknick, M. c060: Extended inference with lasso and elastic-net regularized cox and generalized linear models. J. Stat. Softw. 62, 1–21 (2015).
Google Scholar
Eilers, P. H. C. & Marx, B. D. Flexible smoothing with B -splines and penalties. Stat. Sci. 11, 89–121 (1996).
Article MathSciNet MATH Google Scholar
Therneau, T. M. & Grambsch, P. M. Modeling Survival Data: Extending the Cox Model (Springer New York, 2000) https://doi.org/10.1007/978-1-4757-3294-8.
Book MATH Google Scholar
Heagerty, P. J. & Zheng, Y. Survival model predictive accuracy and ROC curves. Biometrics 61, 92–105 (2005).
Article MathSciNet PubMed MATH Google Scholar

Download references

Acknowledgements

This study was supported by the Klinikum rechts der Isar of the Technical University of Munich and the Else-Kröner-Fresenius Foundation (P27/10//A33/10 to C. Schmaderer). G. Lorenz was supported by a Baxter IIR grant. S. Werfel was supported by a research scholarship from the Klinikum rechts der Isar of the Technical University of Munich.

Funding

Open Access funding enabled and organized by Projekt DEAL.

Author information

These authors contributed equally: Stanislas Werfel and Georg Lorenz.

Authors and Affiliations

Department of Nephrology, Klinikum rechts der Isar, Technical University Munich, Ismaninger Str. 22, 81675, Munich, Germany
Stanislas Werfel, Georg Lorenz, Roman Günthner, Julia Matschkal, Matthias C. Braunisch, Carolin Schaller, Peter Gundel, Stephan Kemmner, Philipp Moog, Uwe Heemann & Christoph Schmaderer
Institute of Medical Informatics, Statistics and Epidemiology, Technical University Munich, Munich, Germany
Bernhard Haller
Transplant Center, University Hospital Munich, Ludwig-Maximilians-University (LMU), Munich, Germany
Stephan Kemmner
Division of Cardiology, Department of Medicine, University of Michigan, Ann Arbor, MI, USA
Salim S. Hayek
Department of Medicine, Rush University Medical Center, Chicago, IL, USA
Christian Nusshag & Jochen Reiser
Departement of Nephrology, University Hospital Heidelberg, Heidelberg, Germany
Christian Nusshag

Authors

Stanislas Werfel
View author publications
You can also search for this author in PubMed Google Scholar
Georg Lorenz
View author publications
You can also search for this author in PubMed Google Scholar
Bernhard Haller
View author publications
You can also search for this author in PubMed Google Scholar
Roman Günthner
View author publications
You can also search for this author in PubMed Google Scholar
Julia Matschkal
View author publications
You can also search for this author in PubMed Google Scholar
Matthias C. Braunisch
View author publications
You can also search for this author in PubMed Google Scholar
Carolin Schaller
View author publications
You can also search for this author in PubMed Google Scholar
Peter Gundel
View author publications
You can also search for this author in PubMed Google Scholar
Stephan Kemmner
View author publications
You can also search for this author in PubMed Google Scholar
Salim S. Hayek
View author publications
You can also search for this author in PubMed Google Scholar
Christian Nusshag
View author publications
You can also search for this author in PubMed Google Scholar
Jochen Reiser
View author publications
You can also search for this author in PubMed Google Scholar
Philipp Moog
View author publications
You can also search for this author in PubMed Google Scholar
Uwe Heemann
View author publications
You can also search for this author in PubMed Google Scholar
Christoph Schmaderer
View author publications
You can also search for this author in PubMed Google Scholar

Contributions

S.W. and C.Schmaderer designed the study. G.L. performed and supervised data collection. R.G., J.M., M.C.B., C.Schaller, P.G., S.K., P.M., U.H., J.R. and C.Schmaderer organized and performed data collection. S.W. performed the statistical and machine learning analyses, made figures and drafted the manuscript. B.H. helped to design and critically assess the statistical and machine learning analyses and draft the manuscript. C.Schmaderer, S.S.H., C.N., J.R., P.M. and U.H. critically revised the manuscript. C.Schmaderer and U.H. acquired funding. All authors gave their final approval.

Corresponding author

Correspondence to Christoph Schmaderer.

Ethics declarations

Competing interests

J. Reiser reports grants from National Institute of Health, Thermo BCT and Nephcure, personal fees from Biomarin, TRISAQ, Astellas, Massachusetts General Hospital, Genentech, Up to Date, Merck, Inceptionsci, GLG, Clearview, outside the submitted work; In addition, J. Reiser reports patents licensed to Miltenyi and TRISAQ, outside the submitted work. J. Reiser reports he is co-founder and share holder of TRISAQ, a bio-pharmaceutical company that develops drugs for kidney disease. All of the remaining authors have nothing to disclose.

Additional information

Publisher's note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary Information

Supplementary Information.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

Reprints and permissions

About this article

Cite this article

Werfel, S., Lorenz, G., Haller, B. et al. Application of regularized regression to identify novel predictors of mortality in a cohort of hemodialysis patients. Sci Rep 11, 9287 (2021). https://doi.org/10.1038/s41598-021-88655-0

Download citation

Received: 27 June 2020
Accepted: 09 April 2021
Published: 29 April 2021
DOI: https://doi.org/10.1038/s41598-021-88655-0
Springer Nature Limited

Application of regularized regression to identify novel predictors of mortality in a cohort of hemodialysis patients

Abstract

Similar content being viewed by others

Dynamic personalized risk prediction in chronic heart failure patients: a longitudinal, clinical investigation of 92 biomarkers (Bio-SHiFT study)

Nomogram for Predicting Cardiovascular Mortality in Incident Peritoneal Dialysis Patients: An Observational Study

Development and validation of a model for predicting the risk of cardiovascular events in maintenance hemodialysis patients

Introduction

Discussion

Methods

Subjects/study population and data collection

Serumbiochemistry and functional studies

Statistical analyses

References

Acknowledgements

Funding

Author information

Authors and Affiliations

Contributions

Corresponding author

Ethics declarations

Competing interests

Additional information

Publisher's note

Supplementary Information

Supplementary Information.

Rights and permissions

About this article

Cite this article

Navigation

Application of regularized regression to identify novel predictors of mortality in a cohort of hemodialysis patients

Abstract

Similar content being viewed by others

Dynamic personalized risk prediction in chronic heart failure patients: a longitudinal, clinical investigation of 92 biomarkers (Bio-SHiFT study)

Nomogram for Predicting Cardiovascular Mortality in Incident Peritoneal Dialysis Patients: An Observational Study

Development and validation of a model for predicting the risk of cardiovascular events in maintenance hemodialysis patients

Introduction

Discussion

Methods

Subjects/study population and data collection

Serumbiochemistry and functional studies

Statistical analyses

References

Acknowledgements

Funding

Author information

Authors and Affiliations

Contributions

Corresponding author

Ethics declarations

Competing interests

Additional information

Publisher's note

Supplementary Information

Supplementary Information.

Rights and permissions

About this article

Cite this article

Share this article

Search

Navigation