Big data-driven establishment and bias comparison of serum ferritin reference intervals in Zhejiang Chinese adults using six algorithms

Abstract

This study systematically compared the performance of six algorithms for establishing serum ferritin (SF) reference intervals (RIs) using data from large-scale health check-up populations (22,359 cases from Hospital A and 555 cases from Hospital B) and 327 anemia patients. Methods included non-parametric (EP28-NP) and parametric (EP28-P) approaches per EP28-A3c guidelines, along with modern algorithms (TMC, refineR, Kosmic, Bhattacharya). Continuous RIs were constructed using GAMLSS. Results showed good consistency in lower RIs limits between EP28-NP and four modern algorithms. Using EP28-NP as an example, male RIs (20–92 years) were 68.8-495.5 ng/mL, while females exhibited age-gradient characteristics: 20–45 years (10.4-132.5 ng/mL), 46–58 years (13.8-241.6 ng/mL), and 59–90 years (43.7-348.8 ng/mL). All algorithms demonstrated satisfactory validation rates (90.91%-98.67%). Notably, application of laboratory-established RIs significantly increased SF abnormality detection rates in anemia patients from 13.2% to 33.6% compared to manufacturer standards. Continuous RIs revealed significant SF concentration increases in males aged 20–29 years and females aged 46–58 years. The study demonstrates that laboratory-developed RIs better reflect population characteristics than manufacturer standards, and recommends that laboratories select appropriate algorithms based on data distribution to establish localized reference criteria.

Introduction

Serum ferritin (SF) is recognized by the World Health Organization as a key biomarker for assessing iron deficiency and iron overload, playing an important role in both physiological and pathological processes¹. Low SF levels are highly specific for iron deficiency anemia. Meanwhile, as an acute-phase reactive protein, SF can rise in various non-overload conditions, such as inflammation, hepatic cell injury, renal disease, infectious diseases, malignancies, and metabolic syndrome².

Reference intervals (RIs) are the fundamental basis for interpreting laboratory results. In China, 91.4% of the SF RIs currently used are derived from manufacturer manual³. For example, the Abbott Alinity i analyzer uses SF RIs defined by 32 healthy males (21.8–276.7 ng/mL) and 60 healthy females (4.6–204.0 ng/mL)⁴. According ISO 15189:2022 and CLSI EP28-A3c, laboratories should establish RIs appropriate for their specific service populations^5,6.

The establishment of RIs is divided into two principal methods: the direct method and the indirect method. Despite the accuracy of direct methods, they are constrained by complicated implementation, difficulty in recruiting truly healthy individuals, and high costs. Consequently, indirect methods have gained popularity. These methods leverage accumulated laboratory data, making RIs determination more feasible, cost-effective, and rapid^7–9. The CLSI guideline EP28-A3c is being updated to EP45-ED1, which will classify Kosmic, TMC, and refineR as second-generation indirect methods¹⁰. Though originally intended for direct methods, The parametric (EP28-P) and non-parametric (EP28-NP) approaches described in EP28-A3c have also found frequent application in indirect RIs research^7,11–15, particularly when the study subjects are healthy populations.

Modern indirect methods include Hoffmann, TMC, refineR, Kosmic, Bhattacharya, and GAMLSS, but their performance characteristics are not widely recognized in many laboratories. TMC involves iterative refinement of initial parameter estimates (λ, µ, σ) under a normality assumption, ultimately deriving RIs from optimized parameters¹⁶. This method has been applied to thyroid hormones and tumor markers^17,18. RefineR employs an inverse modeling approach to isolate the healthy distribution from observed test results and identify the optimal model for RIs construction¹⁹, with applications in hematology, biochemistry, immunology, and thyroid-related hormones^20–22. Kosmic uses Box-Cox transforms and is suitable for non-Gaussian data, truncating intervals within the test result range to estimate the physiological distribution²³. It has been validated for thyroid hormones and pediatric biochemistry^22,24,25. Bhattacharya assumes that a large portion of the mixed data follows a near-Gaussian “healthy” distribution, relying on graphical methods for intuitive RIs extraction^26,27. Applications include biochemistry and total carbon dioxide analysis^25,28. GAMLSS (Generalized Additive Models for Location, Scale, and Shape) is particularly suitable for constructing continuous, age-dependent RIs²⁹.

Previous studies have shown that there are differences in the RIs of SF in terms of gender, age, detection platform, and geographic region, etc. and there is inconsistency in sample size, age distribution, and methods of RIs establishment across different studies^30–32. Male SF levels can be up to four times those of females of the same age, while postmenopausal females can exhibit roughly twice the SF values of premenopausal females. Different analyzers also affect RIs; for instance, Roche and Abbott systems generally yield higher SF measurements than Beckman platforms. In this study, six indirect methods (both classic and modern) were used to analyze large-scale health examination data, thereby establishing and validating sex-specific and age-specific SF RIs on the Abbott Alinity i analyzer.

Materials and methods

Data sources and study cohorts

The study cohorts included establishment, validation, and evaluation groups. The establishment cohort included healthy individuals undergoing routine physical examinations at the First Affiliated Hospital of Zhejiang University School of Medicine (Hospital A) from June 1, 2021 to December 31, 2024. The validation cohort comprised individuals from the First Affiliated Hospital of Zhejiang Chinese Medical University (Hospital B) between January 1, 2023 and December 31, 2024. The evaluation cohort included anemia patients and general patients. The anemia patients were from Hospital A between October and December 2024, and the general patients were from Hospital A in January 2025.

A total of 77,433 candidates from Hospital A were initially identified, all older than 19 years and tested for SF during the study period. Exclusion Criteria: Incomplete demographic data (e.g. missing sex or medical record number; 1,349 cases excluded); Repeated SF measurements from the same individual (only the most recent result retained; 25,828 duplicates excluded); Pregnant women and positive findings on auxiliary examinations (e.g. chest X-ray, ultrasound, CT; 4370 cases excluded). Laboratory test exclusion criteria: Exclusion of individuals with abnormal results for white blood cell count, hemoglobin, mean corpuscular volume (MCV), mean corpuscular hemoglobin concentration (MCHC), alanine aminotransferase (ALT), gamma-glutamyl transferase (GGT), urea, glucose, creatinine, carcinoembryonic antigen (CEA), alpha-fetoprotein (AFP), carbohydrate antigens, body mass index (BMI), or blood pressure outside reference intervals (n = 23,527 excluded). These exclusions aimed to minimize confounding from inflammation, anemia, hepatic/renal dysfunction, malignancies, and metabolic syndromes. After exclusions, 22,359 cases were included in the establishment cohort. The same inclusion/exclusion criteria were applied to Hospital B, yielding 555 subjects for validation.

For the anemia patients in the evaluation cohort, used to assess the abnormality rate of SF RIs in disease states, the inclusion criteria were patients with (male < 120 g/L, female < 110 g/L) who also had SF tested, totaling 327 cases (218 outpatients, 109 inpatients). For the general patients in the evaluation cohort, used to assess the abnormality rate of self-established SF RIs versus Abbott RIs in outpatients and inpatients, the inclusion criteria were outpatients over 19 years old with SF tests (n = 10,622) and inpatients (n = 14,072), totaling 24,694 cases.

Both hospitals are located in eastern China, serving predominantly Han Chinese populations, and are ISO 15,189-accredited. Due to the retrospective nature of the study, the Institutional Review Boards of Hospital A and Hospital B waived the requirement for informed consent. Ethical approval was granted by the respective institutional review boards (Hospital A: IIT2025-0068; Hospital B: 2025-KL-039). All methods were performed in accordance with the relevant guidelines (CLSI EP28-A3c ) and regulations, including the Declaration of Helsinki and its later amendments. The data processing workflow for establishing SF RIs is shown in Fig. 1.

Fig. 1.

Data processing workflow for establishing serum ferritin reference intervals.

Assay methods and analytical performance

Participants fasted for ≥ 12 h prior to venous blood collection. Blood samples (2–3 mL) were drawn using a straight needle into serum separator tubes (VACUTTE^®, Greiner Bio-One, Kremsmünster, Austria), gently inverted 8–10 times, and stored at room temperature. Serum was separated by centrifugation (2000 × g, 10 min) and analyzed within 6 h.

The detection instrument is ARCHITECT Alinity i (Abbott Diagnostics, Ireland), using chemiluminescent microparticle immunoassay with reagent kits (Abbott Diagnostics, Ireland), and the results are traceable to the 1st generation international standard NIBSC 80/602.

During the study period, no changes were made to the environment, instruments, reagents, tubes, etc., and all personnel received uniform training and assessment. Quality control was conducted using Abbott Multichem IA Plus immunocomplex controls (low, medium, and high levels). The long-term cumulative CV% for SF over the past four years was: Low level: 4.36%, Medium level: 3.22%, High level: 3.32%. Monthly calibration was performed using Abbott Diagnostics Ferritin calibrators in a periodic 2-point linear mode. Annual instrument calibration was conducted for the pipetting system, temperature control, and detection system.

Hospital B employed the same instrument, testing procedures, reagents, and QC/calibration protocols, except for differences in reagent lot numbers. Both hospitals have maintained satisfactory performance in external quality assessments (EQA) for the past four years, and method comparison between the two sites showed consistent SF results.

Establishment of reference intervals

Data were extracted from the laboratory information system, including demographics (name, sex, date of birth), test results, instrument identifiers, and testing dates, and stored in .xlsx format. Duplicate entries were removed by retaining only the most recent result per individual (based on medical record number, name, sex, and date of birth). Data were anonymized, and personnel involved in processing signed confidentiality agreements.

For data with significant skewed, Box-Cox transformation was applied to approximate normality or symmetry. Select the λ value that results in the transformed data most closely approximating a normal distribution. The transformation is defined as:

Data exceeding these thresholds were excluded, and the remaining values were back-transformed to the original scale. Tukey’s method was used to detect and remove extreme outliers, defined as observations falling below the LL or above the UL calculated as follows: LL = Q1–1.5 × IQR, UL = Q3 + 1.5 × IQR, where Q1 and Q3 represent the 25th and 75th percentiles, respectively, and IQR=Q3 − Q1. Values outside these limits were excluded prior to subsequent analyses.

Population subgroups were stratified by sex and age according to the Harris and Boyd criteria³³, supplemented by continuous RIs derived via GAMLSS. The final subgroups were defined as: Males: 20–92 years (single age span) ; Females: 20–45 years (premenopausal), 46–58 years (perimenopausal), and 59–90 years (postmenopausal).

EP28-NP: Non-parametric RIs were defined as the 2.5th and 97.5th percentiles after excluding outliers. EP28-P: Following Box-Cox transformation and Tukey’s outlier exclusion, data were back-transformed, and 95% RIs were calculated as mean ± 1.96 standard deviation (SD). Modern Indirect Methods (TMC, refineR, Kosmic): RIs were derived without outlier exclusion. After Box-Cox transformation and the Tukey method, RIs were established using the Bhattacharya. For all methods, 95% RIs (2.5th–97.5th percentiles) with 90% confidence intervals (CI) were established. GAMLSS: Continuous age-dependent RIs for SF were modeled separately for males and females.

Comparison of bias and performance evaluation of indirect methods

A bias ratio (BR) matrix was used to quantify differences in lower and upper RIs limits across algorithms. EP28-NP served as the benchmark for comparisons with the five other indirect methods. refineR was the benchmark for TMC, Kosmic, and Bhattacharya.

The BR for LL and UL was calculated as:

where LL₀ and UL₀ represent the lower and upper limits of the benchmark method. A |BR| of ≥ 0.375 was considered a significant difference^27,34,35.

Using a custom-developed RIs software platform, the performance of six indirect methods was evaluated across five key dimensions: runtime speed, user-friendliness, adaptability to heterogeneous datasets, error frequency, and reliability of results.

Validation and evaluation of reference intervals

Validation was performed using large samples (n ≥ 60 per subgroup). The validation cohort excluded 17 outliers via Tukey’s method. A pass rate ≥ 90% (i.e. ≤10% of validation data outside the RIs) was required for successful validation⁵. Pass rate was calculated as:

Abnormality rates for SF were calculated separately for general patients and anemia patients, stratified by patient category (general vs. anemia), sex, and age subgroups. Abnormality rates were defined as:

Analyses were performed to compare Abnormality rates across subgroups, with results reported as percentages.

Statistical analysis

Data were processed using WPS Office (Kingsoft Corporation, China). Statistical analyses were conducted using SPSS 27.0 (IBM Corporation, USA). Normality was tested with the Kolmogorov-Smirnov (K-S) test. For non-normally distributed continuous variables are reported as median [interquartile range (IQR; P25, P75)]. Between-group differences were evaluated via Mann-Whitney U tests or Kruskal-Wallis tests, with P< 0.05 considered statistically significant.

GAMLSS analyses were performed using the “gamlss” package (version 4.4.0) in R²⁷. UL of age-dependent RIs were analyzed via segmented linear regression. The result distribution plots were generated using the ggplot2 package in R version 4.5.1. Modern indirect methods (TMC¹⁶, refineR¹⁹, Kosmic²³, Bhattacharya²⁶ were implemented in R following published workflows. TMC¹⁶: Truncated minimum chi-square approach for the indirect estimation of reference limits(https://user.math.uni-bremen.de/c05c/TMC); refineR¹⁹: The algorithm can estimate RIs from real-world data consisting of a mixed distribution of non-pathological and pathological test results. (https://cran.r-project.org/web/packages/refineR/index.html); Kosmic²³: Reference Interval Estimation from Mixed Distributions using Truncation Points and the Kolmogorov-Smirnov Distance (https://kosmic.diz.uk-erlangen.de/); Bhattacharya²⁶: The method finds a Gaussian peak in noisy data by focusing on a significant, stable section of the peak that remains largely unaffected by noise (https://sourceforge.net/projects/bellview).

Results

Distribution of serum ferritin across different sex and age groups

Data were grouped according to the Harris and Boyd method (see Fig. 2). Raw and Box-Cox transformed distribution parameters (skewness, kurtosis) for each subgroup are summarized in Table 1. SF exhibited a pronounced right-skewed distribution across all groups, with significantly more high-value outliers than low-value outliers. After Box-Cox transformation, skewness and kurtosis were substantially reduced, bringing each subgroup’s distribution closer to normal.

Fig. 2.

Distribution of serum ferritin data stratified by sex and age.

Table 1.

Serum ferritin distribution (raw and Box-Cox transformed) by sex and age.

Subgroup	Quartiles (ng/mL)	Data type	Skewness	Kurtosis
Male (overall)	230.9 (161.5, 323.6)	Raw data	1.788	6.865
		BOX-COX transformed (λ = 0.4646)	− 0.037	− 0.249
Female (overall)	59.1 (33.8, 97.9)	Raw data	2.391	9.017
		BOX-COX transformed (λ = 0.2626)	− 0.031	− 0.609
Female (20–45 years)	50.8 (30.4, 79.4)	Raw data	1.897	6.19
		BOX-COX transformed (λ = 0.3434)	− 0.044	− 0.633
Female (46–58 years)	86.6 (46.3, 141.0)	Raw data	1.487	2.813
		BOX-COX transformed (λ = 0.5455)	− 0.072	− 0.687
Female (59–90 years)	170.7 (118.3, 229.95)	Raw data	1.412	3.35
		BOX-COX transformed (λ = 0.4646)	− 0.058	− 0.32

Reference intervals for serum ferritin established by indirect methods

The RIs for SF established by six indirect methods are shown in Table 2. SF RIs exhibited substantial inter-individual variability across subgroups, with a 7- to 25-fold difference between lower and upper limits. Overall, male SF levels were significantly higher than those of females, and among females, different age groups exhibited distinct SF ranges. Compared with EP28-NP, EP28-P and Bhattacharya yielded lower lower limits and higher upper limits. In general, RIs derived from the four modern indirect methods were higher than those from EP28-based methods.

Table 2.

Establishment of serum ferritin reference intervals using six indirect methods (ng/mL).

Method	Male (20–92 years) (n = 11,477)	Female (20–45 years) (n = 8377)	Female (46–58 years) (n = 1582)	Female (59–90 years) (n = 923)
EP28-NP	68.8 (45.9–87.1) – 495.5 (454.6–533.5)	10.4 (8.5–13.0) – 132.5 (119.0–144.0)	13.8 (10.5–17.1) – 241.6 (215.9–259.4)	43.7 (33.8–60.1) – 348.8 (326.1–366.9)
EP28-P	65.0 (56.4–74.3) – 500.7 (474.5–527.7)	9.4 (8.7–10.2) – 137.9 (133.4–142.6)	11.9 (8.5–16.2) – 257.0 (229.4–286.6)	44.6 (18.5–80.5) – 351.1 (275.7–434.6)
TMC	73.5 (71.6–75.4) – 518.5 (504.2–532.8)	11.7 (11.0–12.5) – 149.2 (139.4–159.0)	11.1 (8.7–13.4) – 289.2 (248.9–329.4)	42.3 (33.9–50.8) – 342.0 (305.3–378.7)
refineR	75.2 (68.2–89.9) – 548.0 (478.1–589.4)	10.9 (9.2–13.0) – 154.6 (147.2–168.2)	9.8 (4.1–16.2) – 268.6 (214.1–320.0)	34.1 (29.0–89.5) – 327.7 (297.0–378.8)
Kosmic	83.4 (79.6–87.2) – 552.6 (548.8–556.4)	9.5 (8.2–10.8) – 151.9 (150.6–153.2)	6.1 (0.6–11.6) – 261.8 (256.3–267.3)	19.3 (9.9–28.7) – 304.6 (295.2–314.0)
Bhattacharya	60.8 (57.8–63.8) – 502.2 (499.2–505.2)	8.6 (7.6–9.6) – 147.3 (146.3–148.3)	11.0 (6.1–15.9) – 283.6 (278.7–288.5)	40.6 (33.2–48.0) – 362.6 (355.2–370.0)

Note: Data in the table represent LL of the reference interval (90% CI) to UL of the reference interval (90% CI).

Bias ratio and performance evaluation among different indirect methods

BRs for SF RIs established using five indirect methods, with EP28-NP as the benchmark, are presented in Table 3. The results show that EP28-NP and EP28-P were broadly consistent (all BR ≤ 0.265), and the lower limits from TMC, refineR, Kosmic, and Bhattacharya were also largely in agreement with EP28-NP (all BR ≤ 0.313). However, the upper limits generally showed significant deviations (BR ≥ 0.465). Table 4 presents the BR values when refineR is used as the benchmark, comparing it with TMC, Kosmic, and Bhattacharya. Aside from the significantly different upper limit in female patients aged 59–90 years derived from Bhattacharya (BR = 0.466), the others displayed good consistency (|BR| ≤ 0.335).

Table 3.

Bias ratios of serum ferritin reference intervals (EP28-NP vs. 5 indirect methods).

Method	Bias Ratio	Male (20–92 years)	Female (20–45 years)	Female (46–58 years)	Female (59–90 years)
EP28-P	BRLL	-0.035	-0.032	-0.033	0.012
	BRUL	0.048	0.173	0.265	0.030
TMC	BRLL	0.043	0.042	-0.046	-0.018
	BRUL	0.211	0.536	0.819	-0.087
refineR	BRLL	0.059	0.016	-0.0688	-0.1233
	BRUL	0.482	0.710	0.465	-0.271
Kosmic	BRLL	0.134	-0.029	-0.133	-0.313
	BRUL	0.525	0.623	0.348	-0.568
Bhattacharya	BRLL	-0.073	-0.058	-0.048	-0.040
	BRUL	0.062	0.475	0.723	0.177

Table 4.

Bias ratios for serum ferritin reference intervals established by refine R compared to those established by TMC, Kosmic, and Bhattacharya.

Method	Bias ratio	Male (20–92 years)	Female (20–45 years)	Female (46–58 years)	Female (59–90 years)
TMC	BRLL	− 0.012	0.022	0.020	0.109
	BRUL	− 0.215	− 0.147	0.312	0.191
Kosmic	BRLL	0.060	− 0.038	− 0.056	− 0.198
	BRUL	0.034	− 0.074	− 0.103	− 0.308
Bhattacharya	BRLL	− 0.105	− 0.063	0.018	0.087
	BRUL	− 0.335	− 0.199	0.227	0.466

Taking male SF as an example, the performance evaluation of RIs established by six indirect methods is presented in Table 5. The findings suggest that, among the EP28-based approaches, EP28-NP is recommended, whereas among the four modern indirect methods, refineR emerges as the preferred option.

Table 5.

Performance evaluation of six indirect methods for SF RIs establishment.

Evaluation item	EP28-NP	EP28-P	TMC	Kosmic	refineR	Bhattacharya
Run time	27s	4 m 31s	9 m 7s	16 h 30 m	1 m 34s	4 m 10s
R Package Installation	–	–	Complex	Complex	Moderate	Not required
Data Format Conversion	Not required	Not required	Convert to CSV	Convert to TXT	Not required	Convert to TXT
Outlier Exclusion	Manual	Manual	Automatic	Automatic	Automatic	Manual
Normality Transformation	Not required	Manual	Automatic	Automatic	Automatic	Manual
RIs and 90% CI calculation	Manual	Manual	Code-automated	Semi-automated	Code-automated	Semi-automated
Result reliability	Robust	Data normality-dependent	Data normality-dependent	Kernel bandwidth-dependent	Data normality-dependent	Subjective visual judgment bias
Error rate	–	–	Low	Medium	Very low	No errors

Note: Running speed was evaluated on the reference interval software platform; error probability data were derived from the frequency of software errors, interruptions, or unresponsiveness leading to no output during execution across nearly 100 different projects ; “–” indicates “not applicable”.

Validation and evaluation of serum ferritin reference intervals

The validation pass rates of RIs for SF established by six indirect methods are shown in Fig. 3. All methods successfully passed validation, with male pass rates ranging from 91.58% to 96.47% and female pass rates from 90.91% to 98.67%. The majority of failures in validation were located near the lower limit of the RIs.

Fig. 3.

Validation pass rates for serum ferritin reference intervals.

The abnormal rates of SF RIs in general and anemia patients are summarized in Table 6. Compared with the RIs recommended by Abbott, those established via EP28-NP yielded a significant increase in detection of abnormal SF in anemic patients. For instance, the abnormal rate in female patients aged 20–45 years rose from 22.5% (Abbott RIs) to 52.5% (EP28-NP). For male inpatients, the upper-limit abnormal rate dropped from 57.7% to 32.8%, significantly reducing the number of misclassifications for iron overload.

Table 6.

Abnormal serum ferritin rates in general and anemia patients using established RIs (%).

Abnormal rate of RIs	Male (20–92 years)			Female (20–45 years)			Female (46–58 years)			Female (59–90 years)
	Anemia	Outpatient	Inpatient	Anemia	Outpatient	Inpatient	Anemia	Outpatient	Inpatient	Anemia	Outpatient	Inpatient
Number of cases	157	4706	7788	80	1817	1535	48	1910	1594	42	2189	3155
Percentage below Abbott RIs lower limit	8.3	2.8	2.3	22.5	6.1	2.5	25.0	2.0	1.0	0.0	0.1	0.1
Percentage exceeding Abbott RIs upper limit	23.6	44.3	57.7	11.3	4.4	17.9	27.1	20.3	35.1	40.5	41.1	57.8
Percentage below EP28-NP RIs lower limit	29.3	9.7	8.0	52.5	21.4	11.9	33.3	8.6	8.0	14.3	6.4	7.7
Percentage exceeding EP28-NP RIs upper limit	8.9	18.4	32.8	17.5	10.7	25.4	20.8	13.8	29.7	21.4	15.8	35.7

Note: RIs are reference intervals. Abbott RIs: Males, 21.8–276.66 ng/mL; Females, 4.63–204.00 ng/mL. EP28-NP RIs are shown in Table 2.

Continuous age-dependent reference intervals

Continuous changes in SF RIs by sex and age are shown in Fig. 4. SF dataset exhibited pronounced asymmetry. Values below the P₅₀ were tightly clustered, while those above P₅₀ showed wide dispersion. In males, SF RIs have a nonlinear relationship, the P_97.5 percentile increased markedly from 353.8 ng/mL to 518.2 ng/mL (annual increase: 18.27 ng/mL) between 20 and 29 years, followed by two fluctuations in the P₅₀ and higher percentiles during 30–50 years. Beyond age 50, SF levels stabilized with minimal variability, suggesting a plateau in iron storage dynamics during later adulthood .The SF RIs (90% CI) established using EP28-NP for males stratified into 20–29 years and 30–92 years age groups were 69.8 (49.0, 86.6) − 442.4 (407.2, 472.9) ng/mL and 68.6 (45.1, 87.4) − 511.4 (468.7, 552.4) ng/mL, respectively. For females, segmented linear regression of upper RIs revealed distinct phases: (1) Ages 20–45 showed stable, low SF levels (P_97.5 = 175.8 ng/mL; F-test, P = 0.964);(2) Ages 46–58 exhibited a steep linear rise (F-test, P < 0.001) modeled as , increasing from 175.8 ng/mL to 380.2 ng/mL (annual increase: 17.04 ng/mL); (3) Ages 59–80 demonstrated gradual elevation (F-test, P = 0.013) described by .

Fig. 4.

Continuous age-dependent reference intervals for serum ferritin. The shaded areas around the curves represent the estimated 90% confidence intervals. Black boxes.

Discussion

Leveraging real-world laboratory big data and modern indirect algorithms for establishing and evaluating RIs represents a pivotal direction for future RIs research^7,8,31. SF data typically exhibit a right-skewed distribution, and most existing studies employ non-parametric methods to establish SF RIs^12–14. In this study, EP28-NP was used as the benchmark to calculate bias ratios (BRs) with other indirect methods for consistency comparison and evaluation³⁴. As shown in Tables 2 and 3, the RIs derived from CLSI EP28-NP and EP28-P were consistent, while the lower limits of the RIs from four modern indirect methods aligned with EP28 results, but their upper limits were significantly higher. This discrepancy arises because EP28 relies on traditional statistical methods with transparent and interpretable processes, whereas modern indirect methods utilize more flexible, advanced, and intelligent modeling techniques. However, these methods are sensitive to SF’s right-skewed distribution, wide upper-limit variability, sparse high-end data, and occasional outliers, which can lead to inflated upper-limit estimates during Gaussian simulations⁸. Modern indirect methods produced notably higher SF RIs upper limits (e.g. for males: 552.6 ng/mL; for females: 154.6 ng/mL at 20–45 years, 289.2 ng/mL at 46–58 years, and 362.6 ng/mL at 59–90 years). Notably, several studies have reported SF RIs exceeding those established here, likely due to differences in detection platforms or population specificity. The substantial biological variability of SF—12.8% within individuals and 132% between populations³⁶—highlights the necessity of establishing population-specific RIs. Integrating longitudinal laboratory data (e.g. historical SF trends), related biomarkers (e.g. hemoglobin, transferrin saturation), and clinical context (e.g. inflammatory status, liver disease history, iron supplementation) can further enhance diagnostic accuracy.

The four modern indirect methods demonstrated strong consistency (Table 4), aligning with prior literature³⁷, though differences among them may arise from variations in data distributions (e.g. skewness, kurtosis) and the proportion of pathological data. Therefore, project-specific analyses are essential when selecting appropriate indirect methods. Performance evaluations (Table 5) revealed that EP28-NP is the preferred method for establishing SF RIs. However, because EP28-NP lacks automated outlier detection, modern methods more suitable when datasets include mixed populations (e.g. outpatients or hospitalized patients). Method selection should adapt to data characteristics: TMC or refineR is advised for skewed distributions; Kosmic for complex multimodal data; and Bhattacharya quantile decomposition for normally distributed data, although its subjective limitations warrant caution³⁷.

SF RIs exhibit significant sex-specific differences, consistent with previous studies^{12–15,38,39}. Using EP28-NP as an example, the overall RIs were 68.8–495.5 ng/mL for males and 10.8–170.1 ng/mL for females, which are notably higher than those reported in Abbott reagent guidelines⁴. Additionally, as shown in Tables 1 and 2; Fig. 4, adult female SF RIs demonstrate significant age-dependent trends, with a pronounced increase around 50 years of age, consistent with age-related SF RIs patterns established in prior studies using direct³⁹ and indirect methods^12–14. The sex- and age-associated variations in SF may stem from: (1) physiological iron loss differences: menstruation and lactation cause iron depletion in females, and pregnancy further exacerbates iron deficiency; perimenopause reduces iron loss due to irregular menstrual cycles, while postmenopause halts iron loss entirely, leading to increased iron stores and SF levels approaching those of males; (2) Hormonal regulation: Estrogen suppresses hepatic hepcidin synthesis, enhancing intestinal iron absorption and macrophage iron release. Postmenopausal estrogen decline elevates hepcidin levels, reducing intestinal absorption and inhibiting macrophage iron release, ultimately increasing iron storage (e.g. in the liver and spleen) and raising SF concentrations. Conversely, androgens promote iron absorption, utilization, and storage, and can stimulate erythropoiesis, indirectly affecting iron metabolism; (3) Diet and lifestyle: Males typically consume more heme-iron-rich foods (e.g. red meat, animal organs) and consume alcohol more frequently (which boosts iron absorption), while females often eat less or adopt diets that exacerbate lower SF levels. Testosterone stimulates erythropoiesis, indirectly influencing iron metabolism^40,41.

SF levels exhibit regional and ethnic variations^42–44. As shown in Table 2, the lower and upper limits of SF RIs for adults in Zhejiang, China, were significantly higher than manufacturer-recommended RIs, consistent with findings from Chinese studies by Wang¹² and Li et al.⁴⁵, as well as similar observations in Canada⁴⁶. Literature reports indicate that East Asian populations have higher SF levels compared to Europeans, African Americans, and South Asians. These differences may reflect innate biological variations linked to genetic factors, as well as lifestyle and environmental influences such as daily alcohol consumption, body mass index (BMI), and regional dietary patterns^47–51. Due to the unresolved issue of higher-order reference materials for ferritin, manufacturers utilize different International Standard (IS) materials, leading to measurement differences across detection platforms³⁰ and making SF results non-interchangeable. Existing SF studies demonstrate that Abbott and Roche results are comparable but significantly higher than those from Beckman. The SF RIs established in this study are consistent with those established on Abbott and Roche platforms by Wang¹², Rodgers¹³, and Snozek¹⁴, but higher than the RIs established by Ruzhanskaya³⁹ on the Beckman platform.

The clinical decision-making process primarily relies on RIs and clinical decision limits (CDLs)⁵².SF RIs determine whether test results are abnormal. The SF RIs were validated using an external large-scale cohort (Fig. 3) and evaluated for abnormal rates in both general and anemia patients (Table 6). After adopting the self-established SF RIs via EP28-NP, the detection of SF abnormalities in anemia patients improved markedly, demonstrating enhanced sensitivity for identifying iron deficiency anemia. This approach better matches iron metabolism characteristics in elderly populations. Specifically, for female anemia patients, the lower-limit abnormal detection rate increased from 0% to 14.3%, while for female hospitalized patients, the upper-limit abnormal detection rate decreased from 57.8% to 35.7%. These findings provide a more precise laboratory basis for etiological differentiation and personalized diagnosis/treatment of anemia. Moreover, for both outpatients and inpatients, using the self-established SF RIs also noticeably improved abnormal detection rates, more closely reflecting real-world clinical scenarios. SF CDLs are associated with a significantly increased risk of adverse clinical outcomes or can diagnose the presence of specific diseases, requiring effective clinical interventions. However, standardized methods for establishing CDLs are still lacking. In the future, CDLs could be established based on artificial intelligence and big data to further enhance precision diagnosis and treatment capabilities.

Lastly, the application of personalized SF RIs is noteworthy. Certain age subgroups display pronounced boundary effects, as shown in Fig. 4: significant changes occur in males aged 20–29 years and in females aged 46–58 years, with female SF levels being several-fold lower than those of males in the same age ranges. Menstrual blood loss and increased iron utilization during pregnancy contribute to higher intra-individual biological variability in females^53–55 underscoring the necessity of personalized RIs⁵⁶. Regarding the need for stratification in males, the literature reports controversy^12–14,39. In this study, using the Harris and Boyd method, the calculated Z value for the 20–29 years and 30–92 years groups was < 3.0, indicating stratification was not strictly necessary. However, the RIs after stratification and the trends in continuous RIs (Fig. 4) appeared more appropriate. This also suggests that population-based reference intervals may not be suitable for analytes like ferritin, which have large individual variations and are significantly influenced by age, necessitating personalized reference intervals as a supplement^57,58. In clinical practice, continuous RIs can be integrated into the Laboratory Information System to display percentile distributions (P_2.5, P₅, P₁₀, P₂₅, P₅₀, P₇₅, P₉₀, P₉₅, P_97.5) based on gender and age. When a patient’s result deviates from established RIs, color-coded alerts can be provided, offering personalized guidance for clinical decision-making.

This study has several strengths: First, the large volume of data included can reflect the characteristics of real-world population variation. Second, it employed both the classic EP28-A3c and four common new-generation indirect algorithms and compared their biases, providing a basis for the selection of new algorithms. Additionally, it established continuous reference intervals for SF, visually demonstrating trends by sex and age. Finally, it recommends establishing corresponding SF RIs according to different genders and age groups, and conducting external validation and disease evaluation. This study has several limitations. First, it is a single-center study, potentially introducing selection bias, and the established RIs are applicable to the Abbott platform but not necessarily to other platforms. Second, although strict exclusion criteria were employed, the study could not fully account for subclinical inflammation due to the lack of routinely available markers such as CRP, procalcitonin, and ESR in the health examination dataset, nor could it eliminate other potential confounders including disease history, lifestyle habits, medication use, dietary patterns, and iron supplementation, and markers like transferrin, transferrin saturation, and soluble transferrin receptor were not used to definitively confirm the non-iron deficiency status of the subjects. Finally, the study hospital primarily serves adults, so no pediatric data were available.

Conclusion

While significant differences were observed in the upper limits of SF RIs established by EP28-based methods and the four modern indirect methods, these modern methods themselves showed strong mutual consistency. We recommend choosing indirect algorithms based on the specific distribution characteristics of the data. Adult males exhibit substantially higher SF levels compared to females of the same age, and female SF RIs increase progressively with advancing age. Consequently, SF RIs should be stratified by assay methodology, sex, and age to better align with clinical requirements and enhance diagnostic accuracy.

Author contributions

Xinglun Qi: Experiment design, research implementation, data organization, data analysis/interpretation, manuscript writing.Ping Chen: Data collection, data organization, statistical analysis.Yangbin Li: Data collection, data organization, manuscript writing.Lina Fan: Research implementation, data collection, statistical analysis.Dagan Yang : Research supervision, experiment design, data analysis/interpretation, manuscript revision, critical review of intellectual content.

Funding

This work was supported by the National Key R&D Program of China (Grant No. 2022YFC3602302).

Data availability

The datasets and code used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.

Ethical approval

Ethical approval was granted by the respective institutional review boards (Hospital A: IIT2025-0068; Hospital B: 2025-KL-039). All procedures were in accordance with the ethical standards of the institution and with the 1964.