Haematological setpoints are a stable and patient-specific deep phenotype

Abstract

The complete blood count (CBC) is an important screening tool for healthy adults and a common test at periodic exams. However, results are usually interpreted relative to one-size-fits-all reference intervals^1,2, undermining the precision medicine goal to tailor care for patients on the basis of their unique characteristics^3,4. Here we study thousands of diverse patients at an academic medical centre and show that routine CBC indices fluctuate around stable values or setpoints⁵, and setpoints are patient-specific, with the typical healthy adult's nine CBC setpoints distinguishable as a group from those of 98% of other healthy adults, and setpoint differences persist for at least 20 years. Haematological setpoints reflect a deep physiologic phenotype enabling investigation of acquired and genetic determinants of haematological regulation and its variation among healthy adults. Setpoints in apparently healthy adults were associated with significant variation in clinical risk: absolute risk of some common diseases and morbidities varied by more than 2% (heart attack and stroke, diabetes, kidney disease, osteoporosis), and absolute risk of all-cause 10 year mortality varied by more than 5%. Setpoints also define patient-specific reference intervals and personalize the interpretation of subsequent test results. In retrospective analysis, setpoints improved sensitivity and specificity for evaluation of some common conditions including diabetes, kidney disease, thyroid dysfunction, iron deficiency and myeloproliferative neoplasms. This study shows CBC setpoints are sufficiently stable and patient-specific to help realize the promise of precision medicine for healthy adults.

Extended Data Figure 1 |. Characteristics of setpoint coefficients of variation.

a, Mean inter- and intra-patient coefficients of variation (CVs) for each setpoint are shown stratified by subject age. b, Mean intra-patient CV over various time periods. c, Association between setpoint value and setpoint CV. All results were calculated using cohort A. d, Median (+ IQR) absolute difference between setpoints in cohort A estimated using variable numbers of datapoints against setpoints estimated with 20 CBCs (n=1203). Error bars in a-b represent 95% confidence intervals, generated via bootstrapping with 10,000 samples. For results in b, each patient’s CV was calculated using the single time period with the highest number of isolated outpatient CBCs during the study period (2002–2021). Note that results in d do not include MPV, due to its much lower frequency of collection than the other 9 markers.

Extended Data Figure 2 |. Correlations between setpoints and other laboratory tests.

a, Correlation of CBC setpoints with other CBC setpoints and setpoints of other hematologic markers derived from extended research parameters of the Sysmex XN-9000. b, Correlation of setpoints with mean marker values for a wide range of common laboratory tests. c, Correlation of setpoints with common laboratory test marker differences in 10 pairs of demographically matched subjects with differences in a selected setpoint. Results in (c) reflect differences in the marker between matched patients, normalized by the mean and std of the marker difference across the 10 pairs. X-axis labels in panel c reflect the male pairs (M1, M2, …, M5) and female pairs (F1, F2, …, F5) of the cohort. Axis ordering for a-c was performed via hierarchical clustering. Black boxes in a-b refer to regions where absolute correlation coefficients exceed 0.3. Black boxes in c refer to markers for which p > 0.05 in a 2-sided t-test. Full details of correlation calculations and cohort definition are provided in Supplementary Methods. Details of tests in a-c (full names, units, patient data range, etc.) are listed in Supplementary Tables 10–11, and Supplementary Methods.

Extended Data Figure 3 |. Setpoints shift in some pathophysiologic settings.

a, A long-term HCT trajectory in a patient pre- and post-menopause illustrates a shift in the HCT setpoint. b-f, Shifts tend to occur in patient setpoints after pathophysiologic events (c, hypothyroidism, d, splenectomy, e, liver disease) and (f) after pregnancy. g, Effect sizes (mean marker changes illustrated in panels b-f) are similar when using setpoints or a randomly chosen single isolated CBC. h, Precision is higher with setpoints, based on a ratio of p-values from a t-test of effect sizes using setpoints or isolated CBCs, on a log₁₀ scale. Summary characteristics of cohorts in b-f are given in Supplementary Table 9. Lines in b-f reflect unity. Box plots in b-f represent the median (center line), interquartile range (box), with tails extending 1.5 times the interquartile range beyond the box, and all other outliers are plotted as separate points. P-values in b-f were derived from a 1-sample, 2-sided t-test. Exact p-values are 1e-46, 1.5e-18, 0.002, 1.5e-10, and 1e-14 for b-f respectively.

Extended Data Figure 4 |. Heritability of hematologic setpoints.

a-b, Associations of each hematologic setpoint between first-degree relatives (a) and partners (b) in cohorts A-C. c, Heritability estimates of each CBC index derived from studies in the literature^–. Plots in a-b have been age- and gender-corrected via linear regression.

Extended Data Figure 5 |. Manhattan and QQ-plots for each setpoint GWAS analysis.

a, Manhattan plots for the setpoints not shown in Fig. 2. b, Quintile-quintile plots for each setpoint GWAS. Annotations correspond to nearby genes for the primary SNP in each associated locus. Red annotations correspond to novel loci. A full list of significant hits, loci, and nearby genes is given in Supplementary Tables 3–6. P-values in a-b were derived from a linear regression test, and to adjust for multiple comparisons, a significance threshold of 5e-8 was used.

Extended Data Figure 6 |. Comparison of significance and effect sizes between single marker and setpoint GWAS.

a-b, Comparison of effect estimates (a. beta coefficients for SNPS in a GWAS) and corresponding p-values (b) for significant SNPs from GWAS using a single, randomly-chosen isolated CBC or the setpoint. The dashed line in a-b represents unity, and highlighted colors reflect whether either or both markers achieved significance at p<5e-8. For brevity, hits where both p-values were above 5e-8 have been excluded. P-values in a-b were derived from a linear regression test, and to adjust for multiple comparisons, a significance threshold of 5e-8 was used.

Extended Data Figure 7 |. Number of significant loci identified when using averaged CBC values compared to setpoints.

Results are from GWAS analyses following the same quality control measures as the primary analysis but limited to the patients with at least 8 isolated CBCs (N: 19,773). 1, 2, 4, and 8-point averages were taken from a randomly chosen subset of each patient’s isolated CBCs, but limited to the same set of 8 measurements, such that each higher point average contains all the data from the lower point average.

Extended Data Figure 8 |. Setpoint-mortality associations over various time periods and in different cohorts.

a-b, Calculations of setpoint-mortality associations in Cohort B over 2 years (a) and 5 years (b). c, Calculation of setpoint-mortality associations in a less-restricted MGB cohort with setpoint estimates from 2006–2011 (excluding any members of cohorts A-C; n: 50,423), without requirement of no major inpatient stays during the study period. d, Calculation of setpoint-mortality associations in a distinct cohort from the University of Washington Medical Center (UWMC), with setpoints estimated from 2014–2018 (n: 13,864). Error bars in a-d reflect the 95% confidence interval on the mortality rate. All results shown are after exclusion of setpoints outside the reference range (using MGB reference range for a-c, and UWMC reference range for d). Characteristics of the UWMC cohort are given in Supplementary Table 12.

Extended Data 9 |. Validation of associations between setpoints and risk of disease diagnosis at UWMC.

a, Associations between setpoints (estimated between 01-2014–01-2019) and future diagnoses are given for patients from UWMC (n=13,864 patients), based on ICD code analysis. b, Age- and sex-corrected hazard ratios for future disease diagnosis based on a 1-std increase in each setpoint (n=13,864 patients). c, Time from presentation with a pre-diabetic A1c (5.7–6.4%) until a diabetic A1c (>6.4%) stratified by change in MCHC from setpoint (n=2,173 patients). d, Likelihood of an elevated TSH result (>5mIU/L) stratified by presenting MCV and MCV setpoint (n=7,510 patients). e, Likelihood of low ferritin (<10ng/dL) stratified by presenting HGB and HGB setpoint (n=6,285 patients). Each pair of survival curves in a is significantly different (P<1e-5). Error bars in b reflect the 95% confidence interval for the hazard ratio.

Extended Data Figure 10|. Associations of setpoints and presenting lab values with mortality.

1-year mortality rates stratified by setpoint value (estimated from 2002–2006) and worst lab value in 2007. Results show similar stratifications to Fig. 4c. Note that numbers for WBC results may differ slightly from Fig. 4c, due to use of percentile cut-offs instead of specific clinical cut-offs.

Figure 1 |. Hematologic setpoints are stable over decades in states of health.

a, A single healthy patient’s WBC trajectory over 20 years is stable around 6×10³/μL, with occasional transient pathophysiologic disruptions, and a 95% confidence interval (4.5–8.1) about half as wide as the adult reference interval at the study hospital (4.5–11.0). b, Inter- and intra-patient variation in CBC indices over 20 years (long-term) is slightly larger than variation over several weeks (short-term), and both are smaller than inter-patient variation in study cohort A (n=12,407 patients). Short-term intra-patient CVs were derived from the EFLM database. c, The ratio of intra- and inter-patient CVs is below 0.5 for most indices in cohort A. d, Long-term inter- and intra-patient CVs for CBC indices do not vary significantly with sex or self-reported race or ethnicity in cohort A. e, The long-term PLT trajectory distribution for 500 randomly chosen patients is heterogeneous (black lines) and robustly distinguishes some healthy patients, such as those shown with high (yellow), moderate (green), and low (blue) PLT setpoints. Error bars in b-d reflect 95% confidence intervals for the mean, calculated using Z-scores (intra-patient) and via bootstrapping (inter-patient). Stratification of patient CVs by age and over different lengths of time are shown in Extended Data Fig. 1a–b. The full distribution of patient CVs can be seen in Extended Data Fig. 1c. Dotted lines in e reflect the MGH reference interval.

Figure 2 |. Setpoints are a deep phenotype and generate a strong signal for heritability analysis.

a-b, Setpoint and single CBC correlations between partners (a, n=440 patient pairs) are smaller than those for first-degree relatives (b, n=439 patient pairs). c, Heritability estimates for cohorts in a-b derived from setpoints and single CBCs are similar to literature values. d, SNP-heritability estimates from a cohort of 25,254 MGB patients were often higher when using setpoints than when using single CBCs. e, Manhattan plot for a GWAS of HGB setpoints in a cohort of 25,254 MGB patients. f-g. In GWAS using HGB setpoints and single outpatient values, p-values (f) were more frequently significant for setpoints, and effect sizes (g; beta coefficients for each SNP) were similar. h, Setpoints provide greater yield in terms of significant hits than single CBC markers. i, Polygenic score quintiles were only modestly correlated with mean setpoints (and 95% confidence intervals) in a held-out cohort (n=5051). Quintiles for HCT, HGB, and RBC are stratified by sex. Error bars in a-d show 95% confidence intervals. * in d denotes a statistically significant differences (P < 0.05, 2-sided Z-test, exact p-values given in Supplementary File 3). Dashed lines in f-g correspond to unity. Annotations in e correspond to nearby genes for highly significant loci, with red annotations corresponding to novel loci. Raw data plots are shown in Extended Data Fig. 4a–b. Literature heritability estimates are shown in Extended Data Fig. 4c. Equivalent plots of panels e-g for other setpoints are provided in Extended Data Fig. 5–6. Quintile-quintile plots for each setpoint GWAS are shown in Extended Data Fig. 5b. A full list of GWAS hits, association loci, and gene contexts is provided in Supplementary Tables 3–6.

Figure 3 |. Hematologic setpoints are associated with all-cause mortality.

a, 10-year all-cause mortality rate is associated with setpoint quintiles for all CBC setpoints in MGB cohort B (n=14,371 patients) limited to setpoints within the population-wide reference intervals. b-c, Age- and sex-adjusted 10-year mortality hazard ratios are different from 1.0 for most setpoints (b) and setpoint CVs (c) in both MGB Cohort B (n=14,371 patients) and the UWMC validation cohort (n=13,864 patients). Quintiles were calculated separately for males and females. Error bars in a-c show 95% confidence intervals for the mortality rates (a) and mortality hazard ratios (b-c). * in a indicates a significant (p<0.05, chi-squared) difference between the mortality rates of the highest and lowest quintile (exact p-values given in Supplementary File 3). Note that mortality rate for HCT quintile 1 was not significantly different from that of quintile 5 (p=0.075) but was significantly different from quintiles 2, 3, and 4. Results in b-c were normalized to a 1-standard deviation change in the setpoint. Mortality was estimated from the end of the setpoint-estimation period (01-01-2007 for MGB; 01-01-2019 for UWMC). Results for a over different periods of time and validated in multiple separate cohorts are shown in Extended Data Fig. 8. Error bars in all panels reflect 95% confidence intervals.

Figure 4 |. Setpoints are associated with disease diagnosis and may enhance diagnostic accuracy.

a, In Cohort B (n=14,371 patients), an RDW setpoint in the top quartile is associated with subsequent diagnosis of atrial fibrillation, lowest-quartile HCT setpoint with chronic kidney disease, highest-quartile WBC setpoint with type 2 diabetes, lowest-quartile MCHC setpoint with major adverse cardiovascular events (MACE: heart attack, stroke, and heart failure), lowest-quartile RBC setpoint with myelodysplastic syndrome (MDS), and highest-quartile MCV setpoint with osteoporosis. b, Age- and sex-adjusted hazard ratios for diagnosis per 1-standard-deviation increase in setpoint are different from 1.0 in Cohort B and an independent UWMC cohort (n=13,864 patients). c, 1-year mortality risk in Cohort B varies according to WBC setpoint and subsequent WBC result. d, Age- and sex-adjusted mortality hazard ratios for Cohort B were higher using setpoint-based reference intervals (setpoint ± 2*CV; orange) than population-wide reference intervals (blue). e, Risk of progression from early- to late-stage kidney disease was associated with the relationship between the current HCT and its setpoint (7,991 female patients, 6,249 male patients). f, Risk of progression from prediabetes to diabetes was associated the with MCHC and its setpoint (n=4,801 patients), g, low ferritin with HGB and setpoint (n=13,820 outpatients), h, elevated TSH with MCV and setpoint (n=12,383 outpatients), i, JAK2 mutation with PLT setpoint (n=495 patients). Error bars in b, d, and i show 95% confidence intervals. All setpoint percentiles were calculated separately by sex. * in d denote significant difference (P<0.05, logrank test, exact p-values given in Supplementary File 3). Validations of results in a, b, e-h at UWMC are provided in Extended Data Fig. 9. Results for (c) using other CBC indices are given in Extended Data Fig. 10.