Development and validation of a predictive mortality risk score from a European hemodialysis cohort.

Although mortality risk scores for chronic hemodialysis (HD) patients should have an important role in clinical decision-making, those currently available have limited applicability, robustness, and generalizability. Here we applied a modified Framingham Heart Study approach to derive 1- and 2-year all-cause mortality risk scores using a 11,508 European incident HD patient database (AROii) recruited between 2007 and 2009. This scoring model was validated externally using similar-sized Dialysis Outcomes and Practice Patterns Survey (DOPPS) data. For AROii, the observed 1- and 2-year mortality rates were 13.0 (95% confidence interval (CI; 12.3-13.8)) and 11.2 (10.4-12.1)/100 patient years, respectively. Increasing age, low body mass index, history of cardiovascular disease or cancer, and use of a vascular access catheter during baseline were consistent predictors of mortality. Among baseline laboratory markers, hemoglobin, ferritin, C-reactive protein, serum albumin, and creatinine predicted death within 1 and 2 years. When applied to the DOPPS population, the predictive risk score models were highly discriminatory, and generalizability remained high when restricted by incidence/prevalence and geographic location (C-statistics 0.68-0.79). This new model offers improved predictive power over age/comorbidity-based models and also predicted early mortality (C-statistic 0.71). Our new model delivers a robust and reproducible mortality risk score, based on readily available clinical and laboratory data.

Chronic kidney disease (CKD), which has evolved as a global health burden,[1] affects up to 13% of United States (US)[2] and European[3] adults, who suffer a high incidence of comorbidities and an increased mortality risk[4]. Mortality rates in end-stage renal disease patients on chronic HD, relating mainly to cardiovascular complications and infections, remains higher than that of many cancers or heart failure, at up to 19.2 per 100 person-years versus only 1.2 in the general European population.[5]

An improved ability to identify those patients at an increased risk of death appears desirable for several reasons. Thus, identification of high-risk patients may help focus efforts on risk mitigation strategies. In addition, a valid, general, easy-to-use mortality risk score in HD patients could also be used in patient discussions or when scheduling transplants. In health-care economics, such a score may categorize patients in comorbidity-adjusted registries or reimbursement systems, and inform planning. Furthermore, it may also serve as a research tool—homogenizing the case mix entering clinical trials and targeting specific interventions to particular patient subgroups—thus reducing sample sizes without compromising statistical power.

Previously developed risk scores lack applicability, robustness, and generalizability. An early study by Wright,[6] which categorized patients as ‘low', ‘medium', and ‘high' risk on the basis of age and comorbidities, was popularized by Khan[7] who examined the predictive power of this stratification (referred subsequently here as the Wright–Khan mortality index). A scoring system based on prediction model β-coefficients advanced methodologies, allowing objective assessment of contributory factors and their weighted impact.[8] Recent large and complex studies[9, 10, 11, 12, 13, 14, 15] used internal validation that contributes little to generalizability. Generalizability may be further limited by restricted patient populations,[9, 13] geographic locations,[9, 11, 15] small sample sizes,[11] or insufficient variables.[9, 11, 12, 13, 14] The current study therefore aimed to develop, in a large European cohort of incident HD patients, risk scores for 1- and 2-year all-cause mortality and to validate these scores externally in a similarly sized, predominantly prevalent HD population.

RESULTS

Study population

Between 1 January 2007 and 31 December 2009, 11,508 patients were recruited into the second Analyzing Data, Recognizing Excellence and Optimizing Outcomes (ARO) cohort (AROii; Figure 1). Thirty-seven percent of patients initiated HD within Fresenius Medical Care (FME) facilities; nevertheless, the overall median dialysis vintage was only 4 days upon admission. Nonchronic HD patients, those with no laboratory data, and/or those with a history of transplantation (alone or combined; N=773) were excluded. In addition, 1013 patients left the study during baseline, leaving 9722 patients. During the first and second year of follow-up, 1060 (10.9%) and 654 (9.4%) deaths were reported, respectively, giving 1- and 2-year mortality rates of 13.0 (95% CI 12.3–13.8) and 11.2 (95% CI 10.4–12.1) per 100 person-years, respectively. In the first year, 344 (3.5%) patients left the study owing to a renal transplant, and 1338 (13.8%) patients were lost to follow-up (LTFU); in the second year, 288 patients (4.1%) received renal transplants and 600 (8.6%) patients were LTFU. Patients LTFU did not differ greatly from those who were not (Supplementary Table S1 online). Of the 1938 LTFU patients, 527 (27.2%) patients returned to FME after their follow-up stop date. Patients lost or not lost to transplantation are shown in Supplementary Table S2 online.

Figure 1

Derivation of the AROii study population.

Table 1 shows baseline characteristics of the study populations. Although AROii and Dialysis Outcomes Practice Patterns III (DOPPS III) patients were similar in many aspects, we noted some differences. The baseline vascular access differences between AROii and the third Dialysis Outcomes Practice Patterns (DOPPS) cohort patients may be explained by the mix of incident and prevalent patients in DOPPS. Additional differences include geography, dialysis vintage, smoking habits, diabetes, cancer, and cardiovascular disease history. Notably, the proportion of patients dying in each cohort was similar. Within the DOPPS III cohort, mean dialysis vintage differed by ‘region' (Europe: 4.1±5.5 years; Japan: 6.9±7.1 years; North America: 3.4±4.1 years; Australasia: 4.5±5.0 years).

Table 1

Baseline characteristics of the study populations and subpopulations

Parameters	Ma/NMb	AROii (0–3 Mo)c (N=9722)	AROii (3–6 Mo)d (N=8783)	DOPPS III (0–3 Mo)e (N=10,615)
Incident/prevalent on dialysis (%)	NM	100/0	100/0	16/84
Dialysis vintage (months)	NM	0.5±1.1	3.4±1.1	53.9±67.9
Dialysis vintage (months) (median [IQRf])	NM	0.1 [0.0, 0.5]	3.1 [3.0, 3.5]	27.9 [9.3, 70.9]
1-year all-cause deaths	—	1060 (10.9)	910 (10.4)	1048 (9.9)
2-year all-cause deaths	—	1714 (17.6)	1465 (16.7)	1736 (16.4)
Geography:
Europe	—	9722 (100)	8783 (100)	4994 (47)
Japan	—	0	0	2743 (25.8)
North America	—	0	0	2190 (20.6)
Australasia	—	0	0	688 (6.5)
Age at baseline (years)	NM	64.4±14.7	64.3±14.7	63.4±14.3
Genderg	NM
Female		3904 (40.2)	3550 (40.4)	4420 (41.6)
Male		5817 (59.8)	5232 (59.6)	6195 (58.4)
Smoking statusg	NM
Nonsmoker		3608 (37.1)	3294 (37.5)	5078 (47.8)
Former		1659 (17.1)	1506 (17.1)	2552 (24.0)
Current		736 (7.6)	685 (7.8)	1505 (14.2)
Missing		3719 (38.3)	3298 (37.5)	1480 (13.9)
Body mass index (kg/m2)g	M	26.0±7.8	26.0±5.8	25.1±6.1
Missing		683 (7.0)	474 (5.4)	1038 (9.8)
History of diabetes	NM	2740 (28.2)	2536 (28.9)	4189 (39.5)
History of cardiovascular disease	NM	2480 (25.5)	2430 (27.7)	6995 (65.9)
History of cancer	NM	557 (5.7)	504 (5.7)	1347 (12.7)
Chronic kidney disease etiologyg	NM
Hypertension/vascular		1548 (15.9)	1415 (16.1)	1764 (16.6)
Glomerulonephritis		895 (9.2)	811 (9.2)	2415 (22.8)
Diabetes		2335 (24.0)	2136 (24.3)	2901 (27.3)
Tubulo-interstitial		1062 (10.9)	949 (10.8)	968 (9.1)
Polycystic kidney disease		534 (5.5)	490 (5.6)	595 (5.6)
Miscellaneous/other		3106 (31.9)	2793 (31.8)	1220 (11.5)
Invalid/Missing		242 (2.5)	189 (2.2)	752 (7.1)
Vascular access in the first 90 daysg	M
No change: Fistula or graft		3154 (32.4)	3990 (45.4)	7802 (73.5)
No change: Catheter		2908 (29.9)	2062 (23.5)	1637 (15.4)
Change: Fistula/graft to catheter		211 (2.2)	193 (2.2)	99 (0.9)
Change: Catheter to fistula/graft		922 (9.5)	809 (9.2)	355 (3.3)
Other		0	0	158 (1.5)
Missing		2527 (26.0)	1729 (19.7)	564 (5.3)
Actual blood flow (ml/min)g	M	302.7±131.3	327.1±259.7	294.0±90.0
Missing		920 (9.5)	794 (9.0)	925 (8.7)
Dialysis adequacy (equilibrated Kt/V)g	M	1.3±0.5	1.4±0.3	1.5±0.3
Missing		718 (7.4)	458 (5.2)	1375 (13.0)
Intradialytic weight change (kg; post-pre)g	M	−1.7±0.8	−1.9±0.8	−2.2±2.7
Missing		75 (0.8)	36 (0.4)	301 (2.8)
Hemoglobin (g/l)g	M
<100		2796 (28.8)	935 (10.6)	1248 (11.8)
100 to <120		4966 (51.1)	4015 (45.7)	5473 (51.6)
⩾120		1638 (16.8)	3613 (41.1)	3849 (36.3)
Missing		322 (3.3)	220 (2.5)	45 (0.4)
Ferritin (μg/l)g	M
<500		3847 (39.6)	2386 (27.2)	3342 (31.5)
⩾500		4822 (49.6)	5383 (61.3)	6629 (62.4)
Missing		1053 (10.8)	1014 (11.5)	644 (6.1)
C-reactive protein (mg/l)g	M	15.6±25.3	13.3±23.2	13.7±80.6
Missing		2393 (24.6)	2551 (29.0)	4150 (39.1)
Serum albumin (g/l)g	M
<35		2518 (25.9)	1554 (17.7)	2516 (23.7)
⩾35		5948 (61.2)	5964 (67.9)	7713 (72.7)
Missing		1256 (12.9)	1265 (14.4)	386 (3.6)
Cholesterol (mmol/l)g	M
<3.6		1722 (17.7)	1231 (14.0)	2890 (27.2)
3.6 to <6.0		4589 (47.2)	3463 (39.4)	5385 (50.7)
⩾6.0		532 (5.5)	423 (4.8)	360 (3.4)
Missing		2879 (29.6)	3666 (41.7)	1980 (18.7)
LDL-cholesterol (mmol/l)g	M
<2.6		3240 (33.3)	2385 (27.2)	3692 (34.8)
2.6 to <3.3		1281 (13.2)	859 (9.8)	789 (7.4)
3.3 to <4.1		653 (6.7)	467 (5.3)	341 (3.2)
4.1 to <4.9		182 (1.9)	143 (1.6)	82 (0.8)
⩾4.9		87 (0.9)	42 (0.5)	23 (0.2)
Missing		4279 (44.0)	4887 (55.6)	5688 (53.6)
Creatinine (μmol/l)g	M	565.4±187.6	614.1±201.7	777.9±256.4
Missing		926 (9.5)	925 (10.5)	52 (0.5)
Calcium (mmol/l)g	M
<2.1		2742 (28.2)	1884 (21.5)	1576 (14.8)
2.1 to <2.6		6368 (65.5)	6388 (72.7)	8552 (80.6)
⩾2.6		126 (1.3)	133 (1.5)	377 (3.6)
Missing		486 (5.0)	378 (4.3)	110 (1.0)
Phosphate (mmol/l)g	M
<0.8		291 (3.0)	271 (3.1)	83 (0.8)
0.8 to <1.5		4583 (47.1)	3958 (45.1)	3390 (31.9)
⩾1.5		4527 (46.6)	4335 (49.4)	7064 (66.5)
Missing		321 (3.3)	219 (2.5)	78 (0.7)
Parathyroid hormone (ng/l)g	M
< 150		2557 (26.3)	2762 (31.4)	3426 (32.3)
150 to <300		2571 (26.4)	2260 (25.7)	2964 (27.9)
300 to <600		1871 (19.2)	1327 (15.1)	2059 (19.4)
⩾600		731 (7.5)	463 (5.3)	946 (8.9)
Missing		1992 (20.5)	1971 (22.4)	1220 (11.5)

Abbreviations: AROii, second Analyzing Data, Recognizing Excellence and Optimizing Outcomes (ARO) cohort; DOPPS, Dialysis Outcomes and Practice Patterns Survey; LDL, low-density lipoprotein.

Categorical variables are reported using n (%). Continuous variables are reported using mean±s.d.

Factors considered modifiable.

Factors considered non-modifiable.

AROii derivation data set using a 0- to 90-day baseline.

AROii derivation data set using a 90- to 180-day baseline.

DOPPS III validation data set using a 0- to 90-day baseline.

Inter quartile range.

Variables where missing values were imputed.

Predictors of mortality

In our main AROii analysis (based on a first 3-months on follow-up baseline), increasing age, low body mass index, and a cardiovascular disease or cancer history were independently associated with both 1- and 2-year mortality (Table 2). Former or current smokers were at a greater risk within 2 years but not at 1 year, as were patients with a CKD etiology of diabetic nephropathy or tubulo-interstitial disease. Of the dialysis quality parameters, baseline use of, or change to, vascular access via a catheter was associated with an increased risk for both time periods, as was lower actual blood flow.

Table 2

Risk markers for 1- and 2-year all-cause mortality, with associated derived risk score points, in a European incident hemodialysis cohort

	2-year all-cause mortality		1-year all-cause mortality
Parameter (unit) and values	HRa (95% CIb)	Pointsc	HR (95% CI)	Points
Age—continuous (years)	1.04 (1.03–1.04)		1.03 (1.03–1.04)
Age—categorical (years)
⩽39		−5		−5
40 to 49		−2		−2
50 to 59		0		0
60 to 69		2		2
70 to 79		4		4
⩾80		6		6
Smoking status
Nonsmoker	1	0
Former	1.16 (1.03–1.31)	1
Current	1.28 (1.02–1.59)	1
Body mass index (kg/m2)
<18.5	1.66 (1.36–2.03)	3	1.49 (1.15–1.92)	2
18.5 to <25	1	0	1	0
25 to <30	0.88 (0.78–0.99)	−1	0.97 (0.84–1.13)	0
⩾30	0.80 (0.69–0.92)	−1	0.84 (0.70–1.01)	−1
Cardiovascular disease history
Yes	1.30 (1.17–1.44)	1	1.33 (1.17–1.51)	2
No	1	0	1	0
Cancer history
Yes	1.75 (1.49–2.05)	3	1.93 (1.60–2.32)	4
No	1	0	1	0
Chronic kidney disease etiology
Hypertension/vascular	1.00 (0.79–1.27)	0
Glomerulonephritis	1	0
Diabetes	1.43 (1.14–1.79)	2
Tubulo-interstitial	1.30 (1.03–1.66)	1
Polycystic kidney disease	0.83 (0.58–1.20)	−1
Miscellaneous/other	1.12 (0.89–1.39)	1
Vascular access
No change: Fistula/graft	1	0	1	0
No change: Catheter	1.42 (1.25–1.61)	2	1.48 (1.23–1.76)	2
Change: Fistula/graft to catheter	1.32 (1.06–1.63)	2	1.46 (1.13–1.89)	2
Change: Catheter to fistula/graft	1.08 (0.88–1.34)	0	1.12 (0.84–1.50)	1
Actual blood flow (ml/min)
<267.0	1	0	1	0
267.0–<298.7	0.89 (0.77–1.03)	−1	0.86 (0.72–1.03)	−1
298.7–<332.1	0.88 (0.76–1.02)	−1	0.82 (0.69–0.98)	−1
⩾332.1	0.83 (0.71–0.97)	−1	0.82 (0.67–1.00)	−1
Hemoglobin (g/l)
<100	1.24 (1.11–1.38)	1	1.30 (1.14–1.49)	2
100 to <120	1	0	1	0
⩾120	0.82 (0.71–0.96)	−1	0.86 (0.71–1.06)	−1
Ferritin (μg/l)
<500	0.85 (0.76–0.96)	−1	0.84 (0.74–0.97)	−1
⩾500	1	0	1	0
C-reactive protein (mg/l)
<2.6	1	0	1	0
2.6-<7.0	1.31 (1.08–1.59)	2	1.25 (0.97–1.61)	1
7.0-<18.2	1.58 (1.31–1.90)	3	1.63 (1.29–2.05)	3
⩾18.2	2.11 (1.77–2.53)	4	2.21 (1.78–2.75)	5
Serum albumin (g/l)
<35	1.42 (1.27–1.59)	2	1.66 (1.42–1.94)	3
⩾35	1	0	1	0
Creatinine (μmol/l)
<431.1	1.46 (1.23–1.72)	2	1.45 (1.19–1.76)	2
431.1–<539.2	1.19 (1.01–1.41)	1	1.13 (0.91–1.40)	1
539.2–<672.9	1.09 (0.92–1.30)	0	1.09 (0.87–1.35)	0
⩾672.9	1	0	1	0
Calcium (mmol/l)
<2.1			1.11 (0.98–1.27)	1
2.1 to <2.6			1	0
⩾2.6			1.68 (1.06–2.65)	3

Multivariate analysis. Parameters significant at the 5% level shown.

HR, hazard ratio.

CI, confidence interval.

Risk-score points.

Lower hemoglobin concentrations were associated with an increased risk for 1- and 2-year mortality; higher levels were linked with better survival. Baseline inflammation (increased C-reactive protein concentrations and high ferritin levels) was highly predictive of mortality at both 1 and 2 years. Malnutrition and/or inflammation, as evidenced by low concentrations of serum albumin, was also consistently predictive. Predialysis serum creatinine represented an additional risk marker, with lower values associated with higher risk, probably reflecting decreased muscular mass and potentially protein wastage in addition to low serum albumin. Finally, hypercalcemia was associated with a higher 1-year mortality risk.

The results obtained using a 90- to 180-day baseline were remarkably consistent with 0- to 90-day baseline observations, or when LTFU patients were coded as deceased (Supplementary Table S3 online). Of note, the relationship between predialysis serum creatinine and mortality was evident in both analyses, suggesting that any residual renal function at the time of HD initiation in this incident dialysis population could not fully explain this association when a 0- to 90-day baseline was applied.

Risk-score derivation and application

When hazard ratios (HRs) were converted to risk-score points, extreme age had the greatest risk contribution (Table 2). A cancer history was generally more disadvantageous than a cardiovascular disease history. Among laboratory parameters, elevated C-reactive protein concentrations contributed the greatest risk, followed by low albumin and creatinine values. Although lower hemoglobin contributed additive risk, higher hemoglobin values and lower ferritin concentrations contributed most to lowering the risk score.

The risk percentage attributable to risk-score totals differed by follow-up length (Figure 2). The contribution of modifiable risk markers increased as the risk score increased (Supplementary Figure S1 online), but only marginally around 50% of the total risk.

Figure 2

Convenient risk-point calculator printout, including conversion from risk points to estimated all-cause mortality and subsequent categorization in low-, intermediate-, and high-risk groups.

Internal discrimination and calibration

The distribution of 1- and 2-year risk-score points for patients with and without events is shown in Figure 3, with the intersection point between patients—8 and 9 points, respectively—defining ‘high-' and ‘low-'risk patients. On applying these cutoffs, the risk score was highly sensitive (2- and 1-year sensitivity 70.7% (95% CI 68.5–72.8%) and 81.5% (95% CI 79.2–83.9%), respectively) but slightly less specific (2- and 1-year specificity 66.0% (95% CI 65.0–67.0%) and 56.4% (95% CI 55.3–57.4%), respectively; Table 3). By extending this risk categorization to tertile of increasing risk, our risk scores effectively separated patients in real-life clinical terms; the proportion of patients in AROii who actually died within 1 and 2 years increased significantly as tertile of risk increased from ‘low' through ‘medium' to ‘high' (all chi-squared for trend P values <0.001, respectively; Table 4). Calibration curves—which essentially answer the question ‘do close to x of 100 patients with a risk prediction of x% have the outcome?'[16]—demonstrate a strong linear relationship between predicted and actual 1- and 2-year mortality (Figure 4). Greater calibration was observed for 2 years (R2=0.98) than for 1 year (R2=0.94), possibly reflecting fewer events in the latter; the consistently lower predicted versus observed mortality in both accords with the lower specificity described above.

Figure 3

The distribution of 1- and 2-year risk score points for European incident hemodialysis patients with and without the event of interest.

Table 3

Sensitivity and specificity of the mortality risk scores among European hemodialysis patients when applied to high- and low-risk groups

Risk Period	Risk	Died (%)	Survived (%)	Total
2-year	Low	503 (8.7)	5285 (91.3)	5788
	High	1211 (30.8)	2723 (69.2)	3934
	Total	1714	8008	9722
	Sensitivity: 70.7% [95% CI 68.5–72.8%]
	Specificity: 66.0% [95% CI 65.0–67.0%]
1-year	Low	196 (3.9)	4883 (96.1)	5079
	High	864 (18.6)	3779 (81.4)	4643
	Total	1060	8662	9722
	Sensitivity: 81.5% [95% CI 79.2–83.9%]
	Specificity: 56.4% [95% CI 55.3–57.4%]

Abbreviation: CI, confidence interval.

Table 4

Risk stratification capacity of the risk score: estimated versus actual all-cause mortality in hemodialysis cohorts for patients classed as ‘Low', ‘Medium', and ‘High' risk based on their risk score

Dataset	Model-based estimated risk of all-cause mortality	Actual all-cause mortality
		2-year		1-year
		Died (%)	Did not die (%)	Died (%)	Did not die (%)
AROii	Lowa	112 (3.9)	2728 (96.16)	69 (2.2)	3077 (97.8)
	Mediumb	391 (13.3)	2557 (86.7)	284 (8.6)	3009 (91.4)
	Highc	1211 (30.8)	2723 (69.2)	707 (21.5)	2576 (78.5)
DOPPS	Lowd	195 (5.8)	3151 (94.2)	69 (2.4)	2804 (97.6)
	Mediume	537 (15.2)	2990 (84.8)	322 (8.3)	3571 (91.7)
	Highf	1004 (26.8)	2738 (73.2)	657 (17.1)	3192 (82.9)

Abbreviations: AROii, second Analyzing Data, Recognizing Excellence and Optimizing Outcomes (ARO) cohort; DOPPS, Dialysis Outcomes and Practice Patterns Survey.

Lower tertile of risk (<15% for 2 years;<9% for 1 year).

Intermediary tertile of risk (15% to <29% for 2 years; 9% to <19% for 1 year).

Upper tertile of risk (⩾29% for 2 years; ⩾19% for 1 year).

Lower tertile of risk in the DOPPS population (<15% for 2 years;<9% for 1 year).

Intermediary tertile of risk in the DOPPS population (15% to <29% for 2 years; 9% to<19% for 1 year).

Upper tertile of risk in the DOPPS population (⩾29% for 2 years; ⩾19% for 1 year).

Figure 4

The relationship between predicted and observed 1- and 2-year mortality in a European incident hemodialysis cohort.

Risk-score validation

The predictive 1- and 2-year risk scores were highly discriminatory when applied externally to the DOPPS population (Table 5). Although generalizability remained high when the DOPPS population was restricted to distinct geographic locations, small ‘regional' differences were noted, with the predictive value being lower in North America and higher in Japan. Risk stratification capacity was also good, with observed mortality increasing with tertile of increasing predicted risk (chi-squared for trend P values <0.001; Table 4).

Table 5

External validation for the AROii mortality risk score at 1 and 2 years in DOPPS, C-statistic

DOPPS population	C-statistic by mortalitya
	2-year	1-year
Incident on hemodialysis	0.73–0.75	0.75–0.76
Prevalent on hemodialysis	0.72	0.73
All	0.72	0.72–0.73
Europe	0.72	0.73–0.74
North America	0.68–0.69	0.69–0.70
Australia	0.71–0.73	0.72–0.74
Japan	0.77–0.79	0.77–0.79

Range over 10 iterations.

Additional discrimination over existing scores

When the previously published Wright–Khan[6, 7] classification was applied, 3381 (35%), 4248 (44%), and 2093 patients (21%) were classified as low, medium, and high risk, respectively. Compared with medium-risk patients, low-risk patients experienced a lower event rate (HR 0.41; 95% CI 0.36–0.48), whereas high-risk patients experienced a higher rate (HR 1.80; 95% CI 1.63–2.00). In this dialysis population, the predictive power of the Wright–Khan classification was moderate (area under the curve (AUC) 0.66; Table 6). The addition of ARO score predictors improved the predictive power (AUC 0.74), with a net 24 and 27% of patients with and without events, respectively, correctly reclassified. Dialysis and laboratory parameters appeared to have the greatest impact.

Table 6

Additional 2-year all-cause mortality discriminatory ability, conferred by different risk predictors, in a European incident hemodialysis cohort

Analysis	Original model	New model	AUC	Δ AUC	Abs IDI	NRIEvents	NRINon-events
Wright–Khan variable analysis
Independent addition of variables
	Wright–Khan	—	0.661			—	—
	Wright–Khan	+medical history	0.679	0.018	0.005	0.23	−0.07
	Wright–Khan	+ clinical factors	0.677	0.016	0.007	0.08	0.07
	Wright–Khan	+ dialysis	0.687	0.026	0.012	0.33	−0.08
	Wright–Khan	+ labs	0.722	0.061	0.041	0.20	0.24
Cumulative addition of variables
	Wright–Khan	—	0.661			—	—
	Wright–Khan	+medical history	0.679	0.018	0.005	0.23	−0.07
	Wright–Khan+medical history	+ clinical	0.687	0.008	0.006	0.08	0.07
	Wright–Khan+medical history+clinical	+ dialysis	0.701	0.015	0.010	0.27	0.01
	Wright–Khan+medical history+clinical+dialysis	+ labs	0.738	0.036	0.034	0.19	0.21
All vs. Wright–Khan alone
	Wright–Khan	+medical history+clinical+ dialysis+labs	0.738	0.077	0.056	0.24	0.27
Liu variable analysis
Independent addition of variables
	Liu	—	0.601	—	—	—	—
	Liu	+ age	0.696	0.094	0.046	0.35	0.11
	Liu	+medical history	0.610	0.009	0.005	−0.65	0.80
	Liu	+ clinical factors	0.622	0.021	0.007	−0.01	0.15
	Liu	+ dialysis	0.642	0.041	0.013	0.27	0.02
	Liu	+ labs	0.699	0.098	0.051	0.20	0.28
Cumulative addition of variables
	Liu	—	0.601	—	—	—	—
	Liu	+ age	0.696	0.094	0.046	0.35	0.11
	Liu+age	+medical history	0.700	0.004	0.003	−0.62	0.60
	Liu+age+medical history	+ clinical	0.709	0.009	0.007	0.05	0.16
	Liu+age+medical history+ clinical	+ dialysis	0.721	0.012	0.010	0.32	−0.04
	Liu+age+medical history+ clinical+dialysis	+ labs	0.750	0.029	0.030	0.17	0.23
All vs. Liu alone
	Liu	+ age+medical history+ clinical+dialysis+labs	0.750	0.149	0.096	0.35	0.31

Abbreviations: Abs IDI, Absolute Integrated Discrimination Improvement; AUC, area under the curve; NRI, (category-free) net reclassification improvement.

Wright–Khan: patients classified as low, medium, and high risk according to the score of Wright et al.[6] and Khan et al[7]; Liu: patients classified into tertile of increasing risk according to the comorbidity index of Liu et al[12]; Medical history: CKD etiology (Wright–Khan variable analysis only), history of cancer and/or cardiovascular disease; Clinical: Body mass index, smoking status; Dialysis: Vascular access change, actual blood flow; Labs: serum albumin, C-reactive protein, hemoglobin, ferritin, and creatinine. NRIEvents and NRINon-events correspond, respectively, to the proportion of events/nonevents reclassified correctly minus the proportion of events/nonevents reclassified incorrectly. For ΔAUC, and Abs IDI, a positive number corresponds to more events/nonevents being reclassified correctly.

Applying the Liu comorbidity index,[12] 5315 (55%), 1860 (19%), and 2547 (26%) patients were classified as low (0–3 points), medium (4 points), and high (⩾5 points) risk, respectively, and this variable was predictive of mortality (low- vs. medium-risk HR 0.75; 95% CI 0.66–0.85; high vs. medium risk HR 1.55; 95% CI 1.36–1.77). Nevertheless, the addition of the ARO score variables improved the predictive power (AUC from 0.60 to 0.75), and a net 35 and 31% of patients with and without events, respectively, were correctly reclassified. Initially, the addition of age had the greatest effect, with the subsequent addition of medical and clinical history contributing little to correct reclassification. When dialysis and laboratory parameters were added, however, further correct reclassification was observed. An additional analysis, based on the Liu comorbidity index excluding CKD etiology (in their original study,[12] the score was more predictive when this parameter was removed), gave similar findings (Supplementary Table S4 online).

Risk prediction over shorter time periods

The 2-year score was highly predictive of 1-year death (c-index range 0.74–0.75), although less so than the 1-year score. Importantly, in the subset of patients who had not commenced HD (N=4247), it effectively predicted mortality in the first 90 days (c-index=0.71).

DISCUSSION

We describe a sensitive and discriminate mortality risk score developed using a large European cohort of incident HD patients. The model was robust, with similar performances in incident dialysis patients at 0–90 or 90–180 days into chronic treatment. Of note, our population started dialysis in 2007–2009: it reflects the current state of the art in medical therapy. In contrast, the most recent previous mortality risk model study included patients initiating dialysis in 2002–2004.[14]

Our aim was not to develop a risk score dedicated to incident patients on HD, but a versatile mortality risk prediction tool generalizable to the widest possible HD population, including both incident and prevalent dialysis patients. External validation in DOPPS confirmed this, with a high degree of discrimination observed when we validated the score against the incident subset and the predominantly prevalent component in DOPPS (Table 5). Generalizability to HD in other geographic areas was also apparent. The observed C-statistic generated (∼0.73), although ‘acceptable' rather than ‘excellent,'[17] was comparable with the previous internally validated studies of Couchoud (0.70),[9] Cohen (AUC 0.77),[11] and van Walraven (0.75),[13] as well as in internal validation of the Framingham Risk Score (0.79[18]). The development of a mortality risk score in a large international database such as AROii, with external validation in another independent worldwide data set as DOPPS, goes significantly beyond previous risk prediction tools. Furthermore, we demonstrate that the use of simple clinical, dialysis, and laboratory routine parameters improved predictive ability over more parsimonious models based on comorbidities alone or age and comorbidities.

Five aspects particularly distinguish this from previously developed risk scores. First, prior attempts were based on often-small patient populations confined to one geographic area.[8, 9, 11, 12, 14, 19] Second, in contrast to other studies,[9, 10, 11, 12, 13, 14] we focused exclusively on incident HD patients, thus minimizing survival bias. Third, we studied patients from various countries, socioeconomic groups, and health-care systems, and prospectively collected data without exclusions. Other recent scores have focused on older dialysis patients,[9] transplant wait-listed HD patients,[13] or particular socioeconomic groups.[12] Fourth, we demonstrate the improved discrimination of our risk predictors over older models[6, 7] comprising age and/or comorbidities alone, and reinforce the clinical meaningfulness of our score through risk stratification capacity analyses. Age and comorbidities should both be integrated in a risk prediction tool as main drivers for mortality;[20] retaining only comorbidities may limit predictive power.[12, 21] This is apparent in the current study, in which the addition of age alone to the Liu comorbidity index correctly reclassified a net 35 and 11% of patients with and without events (Table 6). Laboratory parameters provided the most discriminatory advantage in our score, in line with previous observations.[22, 23] Finally, and potentially most relevant clinically, we apply our score to a population selected for HD but who had not yet initiated HD, and effectively predict early mortality in this group.

From a methodological viewpoint, our approach lies toward the simplistic end of the analytic spectrum. Advantageously, the methods are easily replicable and a simple risk calculator could be implemented easily (e.g., in smartphone applications). Like others,[9, 13, 14] we used imputation to deal with missing data; our choice of Cox regression is also customary.[12, 13, 14] Other approaches range from simplistic Kaplan–Meier plots[6] to complicated fractional polynomial[13] or bootstrapped logistic regression models.[9] More complicated models may improve prognostic ability,[24] but additional computational complexity may well outweigh potential benefits.

Generic indices such as Charlson Comorbidity Index (CCI)[25] or the Index of Co-Existent Disease (ICED),[21] although applied to dialysis patients, were not designed for bedside use and required adaptation.[26, 27] Although such scoring systems may be useful in administrative and economical decision-making, they appear to perform less well for mortality risk prediction in HD populations.[20] Our score, in which ∼50% of the variables were potentially modifiable, may also lay the ground for specific intervention studies. In contrast to others,[10] our score includes notably few comorbidities. Another recent study supplemented data on four variables (age, dementia, peripheral vascular disease, albumin) with a ‘surprise' question (‘Would I be surprised if this patient died within the next 6 months?').[11] However, only 6-month mortality and 500 patients derived from five US dialysis centers were assessed. Performance of the ‘surprise' question in other countries, especially on long-term outcomes, is unknown and so far not externally validated.

Clinically, a mortality risk score should be based on routinely measured parameters, which are easy to derive and calculate. For example, the ICED evaluation can take trained people up to 1 h to complete.[28] It should also be based on accurate, objectively measured variables. This excludes, for example, dementia and congestive heart failure (difficult to define in dialysis patients), or subjective parameters such as self-rated health.[29] This practical aspect must be balanced against potential uncontrolled confounding. Highly subjective parameters—such as the ‘surprise' question,[11, 30] however complementary—will invariably depend on physician background/training and patient knowledge. Although we do not advocate for a binary measure of risk, we would argue that the observed high sensitivity (the ability of the score to correctly identify high-risk patients) is clinically advantageous in the dialysis setting, as the false positive rate (those considered low-risk based on their score but who died nonetheless) will be low.

Our approach has limitations. First, our study is based on data generated from a single commercial dialysis provider, and therefore it could be considered less generalizable to the wider HD population. Our risk score performed favorably when applied externally to the DOPPS population, however, suggesting that it can be applied to a large community of HD patients, although we acknowledge that it may be less generalizable to peritoneal dialysis patients. Focusing on one large provider allowed us a clinical database, which indicates that recorded outcomes will reflect patients' diagnoses rather than health-care providers' claims for reimbursement, as might be observed in administrative databases.[10, 12]

Second, we only assessed 90-day, 1-year, and 2-year mortality, whereas 3-[14] or 5-year[24] mortality may also be important. However, where reported, a remarkably similar C-statistic of 0.75 was obtained in longer analyses.[14]

Third, patients with no laboratory data were excluded, but the loss of 500 of ∼11,000 patients in AROii contrasts favorably with a recent registry analysis[14] in which ∼5500 of 11,000 patients were excluded owing to missing data.

Fourth, a relevant portion of patients were LTFU, especially in the first year, partly reflecting our stringent definition. When we assumed that nonreturning patients had died, however, this had no major bearing on our findings.

Fifth, comorbidity severity was not considered. However, this often introduces subjectivity, and including severity grade did not improve the model in another analysis.[26]

Sixth, potential predictors of death particularly in elderly dialysis patients—such as late referral, dependency for transfers, severe behavioral disorders, health-related quality of life, frailty assessment, and unplanned dialysis[9, 31]—could not be assessed in our analysis of routinely captured data. Other predictive parameters in dialysis patients[32, 33, 34, 35] might conceivably improve our score.

Seventh, the inclusion of patients receiving kidney transplants during follow-up may have selected a healthy cohort, as transplant-listed patients tend to be younger and healthier. By treating these events as censored observations, however, we were in accordance with the analytical recommendations of a recent study focusing on the issue of renal transplantation as a competing event in survival analysis in nephrology.[36] We acknowledge, however, that our score may be less generalizable to HD populations with excessively higher transplant rates than ours (∼8%).

Finally, although country-specific predictions might provide further insights, many subgroups would be too small for meaningful analyses. Of note, our score yielded slightly lower C-statistics in the US DOPPS patients compared with patients in Europe and, in particular, with patients in Japan. This suggests potentially unmeasured confounding in the US analysis. Within Europe, the almost identical score performance in the AROii and DOPPS cohort strongly suggests that the AROii population is representative for a larger European population.

In conclusion, we describe a novel mortality risk score, potentially applicable to all incident or prevalent HD patients, with improved predictive power over age-/comorbidity-based models. Such a tool is now available for research purposes, allowing either correction for imbalanced mortality risk within groups or the selection of high-risk patients; the data may also prove to be useful when communicating with patients, for example, through highly visual heat-maps (Supplementary Figure S2 online). Although risk scores may help generate hypothesis, they can only describe associations and not establish causality.

Materials and Methods

The ARO research initiative[37] comprises European FME HD patients. Anonymized patient-level medical history data, plus longitudinal laboratory and medication data are captured quarterly, as are ICD-10-coded hospitalizations and deaths. The study is based on AROii, which includes consecutive incident (<6 months on dialysis) adult patients without renal transplantation history collected from >300 FME facilities in 14 European countries in 2007–2009. Data on chronic dialysis patients (⩾10 contiguous HD sessions) with available laboratory data were restricted further to patients remaining in the study for ⩾3 months (Figure 1). Patient data were anonymized, and informed consent was obtained from all patients by FME.[37] Local ethics committees' approvals were collected.

Follow-up

Follow-up commenced on the date of patients' first FME HD session; time at risk was accrued from the end of baseline (first 3 months of follow-up) until patients experienced the event of interest or were censored: undergoing a renal transplant, being LTFU (>45 days without continuous FME dialysis treatment; LTFU), or end of follow-up (30 September 2011).

Statistical analysis

Statistical analyses, reproduced independently by a second biostatistician, were performed using SAS (Windows version 9.2; SAS, Cary, NC, USA). Baseline demographic, clinical, dialysis, and laboratory data were categorized into quartiles or biologically relevant groups (Table 1). Continuous variables were described using means and standard deviations; categorical data were reported as counts and frequencies. Rates per 100 person-years with accompanying 95% CIs were calculated.

Risk-score derivation

Risk scores were derived using a modified Framingham Heart Study[38] approach. The predictive effect of exposures was determined using baseline Cox regression models. Ten imputed data sets were created using Monte-Carlo Markov Chain multiple imputation to deal with missing observations, with all covariates described in Table 1 fitted initially. Generated coefficients were combined using Rubin's rules,[39] and variables with at least one significant (5% level) stratum were retained. Coefficients were converted to ‘points', with 1 point indicating the risk equivalent of 5 years additional age; a total score was computed for each patient by summing all points. Estimated absolute risk was calculated by substituting the points total for each risk category in the original Cox model. An algorithm describing risk score point derivation in detail and including 1- and 2-year risk equations is included in the Supplementary material online.

A number of analyses were performed to quantify the relationship between risk score predictions and actual mortality. Traditional measures of discrimination—sensitivity and specificity—were calculated, with the predicted low-/high-risk cutoff determined graphically (Figure 3). This discrimination was extended to further assess the risk stratification capacity: patients were categorized as ‘low', ‘medium', or ‘high' risk on the basis of tertile, and this was compared with their actual 1- and 2-year death status. Changes in the proportion of deaths across these strata were assessed using the chi-squared test for trend. Finally, calibration was assessed graphically: patients' predicted survival was binned into 10% risk groups, and these data were plotted against observed death, calculated as the proportion of patients in each group who actually died.

Contribution of modifiable and non-modifiable risk markers

Further analyses were performed to demonstrate potential clinical utility. By choosing, for illustrative purposes, modifiable and non-modifiable risk markers, a risk ‘heat-map' was constructed. Risk markers were stratified in a nested 2 × 2 arrangement, and the product of the points contributing to strata was calculated. A low-/medium-/high-risk ‘traffic-light' color scheme was applied on the basis of tertiles of all calculated points.

By assigning risk scores to patient-level risk markers, the relative contribution of potentially modifiable and non-modifiable (Table 1) risk markers was calculated: if a patient had a score of 15 and 6 of these points were from modifiable factors, then it was assumed that 40% of their risk was modifiable. By applying these proportions to the actual risk percentage of death, the relative contribution of modifiable with increasing risk score was estimated.

External validation was performed using the third DOPPS[19, 40, 41] cohort. This was chosen for reasons of contemporaneousness (DOPPS: 2005–2008; AROii: 2007–2009) and data comparability. The DOPPS study design dictates randomly selected incident and prevalent HD patients from randomly selected dialysis units within each country, with chosen facilities representing the different facilities and regions within each country. The seven-group ICD-10-based CKD etiology category defined in AROii was applied in DOPPS using ICD-9 diagnostic codes. DOPPS data were also limited to patients with ≥3 months' follow-up. The discriminatory ability of the predictive risk-score model was assessed using Harrell's C-statistic.[42] These scores range from 0 to 1, with 1 indicating a perfect prediction and 0.5 indicating a chance prediction; values of 0.7–0.8 could be considered acceptable and those of 0.8–0.9 could be considered excellent.[17]

Sensitivity analyses

Additional analyses were conducted to assess the impact of our methodological choices. It was possible that a proportion of patients LTFU were lost because they withdrew from dialysis as their condition deteriorated, with these patients censored instead of counted as having a mortality event. To assess the potential impact of such a misclassification, we considered all nonreturning LTFU patients as deceased on their last observation date and repeated the analysis. In addition, to test whether the first 90 days of HD, which represents a high-risk period,[43] was less predictive of longer-term mortality, the AROii baseline was reset to 90–180 days and the predictor analysis was repeated.

Additional discrimination

Analyses were conducted to assess the additional discriminatory power of our 2-year risk score over existing risk scores. Those described by Wright–Khan et al.[6, 7] and Liu et al.[12] were chosen to represent European and North American patients, respectively. Patients were classified according to each score and grouped into low, medium, and high risk (implicit in the Wright–Khan paper; tertile of patients' points total (including points for CKD etiology) for the Liu paper). Separately, and starting with a model containing this variable, we added independently and sequentially the medical history, clinical, dialysis, and laboratory parameters. At each stage, we calculated the AUC, the Integrated Discrimination Improvement, IDI (‘the increment in the predicted probabilities for the subset experiencing an event and the decrement for the subset not experiencing an event'), and the category-free NRIevents and NRInon-events (the ‘net' proportion of individuals with/without events reclassified correctly using the new model over the original model'), as proposed by Pencina et al.,[44] and as calculated by Kennedy.[45] [46] For the Liu analysis, the inclusion of age was also assessed, as this comorbidity index omits age.

The discriminatory ability of the 2-year risk-score model over shorter time periods was assessed using Harrell's C-statistic. First, the 2-year score was applied to 1-year mortality. Subsequently, the ability to predict early death (in the first 90 days of dialysis) was assessed using predialysis data for the subpopulation of patients who initiated HD in FME facilities. Where data for any risk-score parameter was missing (e.g., 100% for actual blood flow), the neutral risk-score point was applied—i.e., zero.