Haematological pre-staging of colorectal cancer: Longitudinal trends identify high-risk metastatic phenotypes 24 months prior to diagnosis

Highlights

• •Pre-diagnostic blood trajectories identify high-risk metastatic CRC phenotypes.
• •Progressive anaemia and inflammation may precede synchronous metastatic presentation.
• •Dynamic haematological trajectories distinguish metastatic from non-metastatic CRC.
• •Haematological pre-staging may support risk-adapted imaging and surveillance.

Abstract

Background

The latent systemic phase of colorectal cancer (CRC) offers a critical but often missed window for early intervention. This study evaluates whether longitudinal Complete Blood Count (CBC) trajectories can stratify metastatic risk before diagnosis.

Methods

A retrospective cohort of CRC patients from the Health Department of La Ribera (Spain) was analysed using serial CBC data collected up to 24 months prior to diagnosis. Linear Mixed-Effects Models characterized patient-specific longitudinal trajectories, capturing intra-individual variability over time. In parallel, supervised machine learning models (Random Forest) were applied to evaluate the discriminative capacity of these haematological variables to distinguish synchronous metastatic phenotypes from non-metastatic cases.

Results

Consistent haematological deviations were detected at least 24 months before diagnosis. A distinct "metastatic gradient" emerged, where patients presenting with synchronous metastases exhibited significantly accelerated trajectories, specifically steeper declines in haemoglobin and sharper rises in inflammatory indices (neutrophil and platelet-to-lymphocyte ratios), compared to non-metastatic cases. These alterations were statistically significant even within clinically normal reference ranges. The predictive models achieved effective risk stratification, demonstrating a robust negative predictive value capable of excluding low-risk phenotypes.

Conclusions

Longitudinal CBC monitoring reveals early systemic "red flags" that anticipate aggressive metastatic behaviour, supporting its use as a haematological pre-staging approach. Assessing the velocity of haematological change, could transform routine historical into a cost-effective resource for risk-adapted imaging and personalized surveillance. However, these findings represent a retrospective proof-of-concept; the approach remains investigational and should not be clinically implemented until validated prospectively in independent external cohorts.

Introduction

Colorectal cancer (CRC) remains a critical global health burden, ranking third in incidence and second in mortality [1]. Its prevalence closely parallels the Human Development Index and modifiable lifestyle factors [[2], [3]]. Despite global efforts, alarmingly, early-onset CRC is rising and frequently presents with advanced disease [4,5]. The prognosis is starkly determined by stage: survival drops from 93% in Stage I to approximately 10% in Stage IV, identifying late presentation as the primary driver of mortality [6]. The liver constitutes the predominant site of metastasis [7]. While 20–25% of patients present with synchronous stage IV disease [8], up to 60% will eventually develop liver metastases (LM) [9]. Despite therapeutic progress, the window for curative intervention remains narrow; 70–80% of metastatic patients present with unresectable disease, precluding curative intent [10,11].

The management of colorectal liver metastases (CRLM) has evolved from simple resection to complex multimodal strategies. Significant milestones, including preoperative optimization, downstaging concepts, and extended surgical criteria, have expanded resectability [9,12]. These options have been further broadened by locoregional therapies and transplant protocols [13,14]. Furthermore, precision oncology now guides decision-making through molecular biomarkers (KRAS, MSI, BRAF) [15,16]. Despite all this technical sophistication, most patients are diagnosed when these options can no longer be applied. In contemporary oncology, the focus is frequently placed on reactive "down-staging"—utilising neoadjuvant therapies to reduce an already established, heavily advanced tumour burden. This paradox highlights that refining surgical technique is insufficient; there is an urgent clinical need to shift the focus towards "haematological pre-staging", a proactive observation of the subclinical phase to detect silent systemic progression before conventional staging.

Although screening programmes have improved secondary prevention, significant challenges persist, particularly due to sub-optimal participation and the rising incidence in younger cohorts which are generally excluded from routine screening [5,17]. Current strategies rely on a dichotomy between standard non-invasive screening tools, such as fecal occult blood testing (FOBT) or fecal immunochemical testing (FIT), and definitive but invasive diagnostic procedures, primarily colonoscopy [18,19]. While emerging technologies, including liquid biopsy assays targeting circulating tumour DNA (ctDNA), are promising, they currently lack the validation and cost-effectiveness required for mass implementation [20]. Consequently, there is a critical gap for intermediate risk stratification tools. The clinical priority is to identify scalable, accessible biomarkers capable of flagging high-risk phenotypes among patients with suspected or confirmed disease, thereby optimizing the allocation of invasive and high-resolution diagnostic resources [21,22].

CRC progression is not merely a local event but drives a measurable systemic signature reflecting the host’s changing immune status. Initially, the host relies on effective immunosurveillance mediated by CD8+/Th1 lymphocytes to suppress tumour growth [23,24]. However, as the tumour advances, the host undergoes a profound immune shift, transitioning toward a chronic inflammatory, pro-tumourigenic environment dominated by neutrophils and platelets.

In this advanced phase, pro-tumoural N2 neutrophils facilitate immunosuppression and invasion via cytokines and matrix metalloproteinases (MMPs) [25]. Furthermore, they drive the formation of neutrophil extracellular traps (NETs), which entrap circulating tumour cells and actively prepare the pre-metastatic niche by increasing vascular permeability[26,27]. Concurrently, platelets act as crucial partners in dissemination, protecting tumour cells from immune destruction and releasing growth factors that foster angiogenesis and colonization [28,29]. Collectively, this dynamic remodelling creates a discernible haematological footprint. Yet, because this process evolves gradually, they elude single-time-point testing and can only be appreciated through longitudinal follow-up that traces the course of the disease.

CBC-based models have emerged as promising tools for early CRC detection [[30], [31], [32]]. Markers such as haemoglobin levels, platelet counts, and inflammatory indices -including the neutrophil-to-lymphocyte ratio (NLR) and platelet-to-lymphocyte ratio (PLR)- act as effective predictors of risk [32]. Algorithms like ColonFlag (AUC 0.82) have demonstrated the potential to identify CRC presence 6–24 months prior to clinical diagnosis [30,31,33]. Furthermore, models tracking longitudinal inflammatory markers (e.g. Systemic Immune-Inflammation Index, SII) have enhanced detection capabilities [34,35]. However, a critical limitation remains: most existing models function as binary classifiers ("cancer vs. no cancer") based on static or short-term data. They often fail to capture the trajectory of aggressiveness. While other approaches like metabolomics offer high accuracy, they lack the routine applicability and cost-effectiveness of CBC data [36]. Therefore, there is an unmet need for models that not only detect disease but distinguish its evolutionary stage through the analysis of dynamic trajectories.

We hypothesize that CRC progression is preceded by distinct longitudinal haematological trails acting as a “biological chronometer” of systemic disease. Specifically, we posit that the rate of change in haemoglobin, immune cells, and platelets differs significantly between metastatic and non-metastatic patients long before diagnosis. Based on this hypothesis, the primary aim of this study is to retrospectively evaluate the pre-diagnostic haematological kinetics in a cohort of patients eventually diagnosed with CRC. By tracing historical CBC data backwards up to 24 months prior to diagnosis, we aim to determine whether specific longitudinal trajectories can reliably distinguish those who present with synchronous metastases from those who do not. Integrating Linear Mixed Models with Machine Learning, this study explores whether longitudinal CBC monitoring can function as a “haematological pre-staging” tool to stratify risk and anticipate tumour aggressiveness.

Materials and methods

Study design

A retrospective observational study was conducted at La Ribera University Hospital (Valencian Community, Spain), the reference centre for a population of approximately 250,000 inhabitants. Patients diagnosed with CRC between January 2012 and December 2022 were identified from Electronic Health Records (EHR) and the Minimum Basic Data Set (MBDS) using specific, standardized ICD-9 and ICD-10 diagnostic codes covering primary anatomical locations of colorectal malignancies and secondary metastatic sites (Table 1). To isolate the pre-diagnostic biological signal of synchronous metastasis and prevent unmeasured heterogeneity, metachronous cases (diagnosed >3 months post-primary) were identified and excluded from this initial proof-of-concept analysis for two primary reasons. Firstly, their pre-diagnostic evolution up to the point of primary diagnosis was statistically indistinguishable from the non-metastatic cohort, leading us to hypothesise that the development of metachronous disease is driven by distinct biological patterns that warrant an independent, dedicated study. Secondly, from a clinical standpoint, these patients enter a rigorous follow-up and surveillance protocol immediately following their primary diagnosis, meaning their subsequent metastatic presentation is captured through entirely different clinical pathways (Fig. 1).

Table 1.

ICD-9 and ICD-10 Diagnostic Codes Used for Colorectal Cancer Case Identification and Metastatic Site Stratification.

ICD-9 Code	ICD-10 Code	Clinical Description (ICD-10 Standard)	Type
153	C18.3	MN of hepatic flexure	PS
153.1	C18.4	MN of transverse colon	PS
153.2	C18.6	MN of descending colon	PS
153.3	C18.7	MN of sigmoid colon	PS
153.4	C18.0	MN of caecum	PS
153.5	C18.1	MN of appendix	PS
153.6	C18.2	MN of ascending colon	PS
153.7	C18.5	MN of splenic flexure	PS
153.8	C18.8	MN of overlapping sites of colon	PS
153.9	C18.9	MN of colon, unspecified	PS
154	C19	MN of rectosigmoid junction	PS
154.1	C20	MN of rectum	PS
—	C78.0	SMN of lung	MS
197	C78.00	SMN of unspecified lung	MS
197.1	C78.1	SMN of mediastinum	MS
197.2	C78.2	SMN of pleura	MS
—	C78.3	SMN of other and unspecified respiratory organs	MS
197.4	C78.4	SMN of small intestine	MS
197.5	C78.5	SMN of large intestine and rectum	MS
197.6	C78.6	SMN of retroperitoneum and peritoneum	MS
197.7	C78.7	SMN of liver and intrahepatic bile ducts	MS
—	C78.8	SMN of other and unspecified digestive organs	MS
—	C79.0	SMN of kidney and renal pelvis	MS
—	C79.1	SMN of bladder and other and unspecified urinary organs	MS
198.2	C79.2	SMN of skin	MS
—	C79.3	SMN of brain and cerebral meninges	MS
—	C79.4	SMN of other and unspecified parts of nervous system	MS
—	C79.5	SMN of bone and bone marrow	MS
—	C79.6	SMN of ovary	MS
—	C79.7	SMN of adrenal gland	MS
—	C79.8	SMN of other specified sites	MS

Note: MN = Malignant Neoplasm; SMN = Secondary Malignant Neoplasm; PS = Primary Site; MS = Metastatic Site. Blanks (—) indicate diagnostic conditions where cross-sectional historical extraction was mapped exclusively using ICD-10 electronic system updates, as these specific sub-categories lacked a direct, corresponding equivalent in the legacy ICD-9 coding framework.

Fig. 1.

Study design and analytical workflow. A retrospective cohort of 2551 colorectal cancer (CRC) patients (2012–2022) was stratified into non-metastatic (G.0), hepatic (G.Ia), and extrahepatic (G.Ib) phenotypes. Complete blood count (CBC) data from the 24 months preceding diagnosis were temporally aligned and reorganized into four density-balanced windows. Individual trajectories were modeled using linear mixed-effects models, and trajectory-derived features were evaluated by Random Forest classification.

The study population was classified into three primary phenotypes: G.0 (non-metastatic CRC), G.Ia (CRC with synchronous or early hepatic metastases), and G.Ib (CRC with synchronous or early extrahepatic metastases). Synchronous metastases were defined as lesions detected at the time of diagnosis or within the subsequent three months. This 3-month window accounts for occult metastases that, while biologically active, reach radiological detectability shortly after diagnosis. This rationale is supported by the literature on the tumour doubling time (TDT) of colorectal cancer metastases, which estimates average growth intervals of 2 to 4 months [[37], [38], [39]].

Inclusion and exclusion criteria

Inclusion criteria required a confirmed diagnosis of colorectal cancer (CRC) and the availability of valid complete blood count (CBC) data within the 24 months preceding diagnosis. Exclusion criteria included: (1) incomplete or missing CBC data matrices, which were systematically eliminated to maintain a clean, fully paired dataset and guarantee the integrity of the multivariate classification; (2) patients with documented alternative causes of severe anaemia or systemic inflammation (e.g., giant hiatal hernias with chronic bleeding, advanced chronic renal failure, recent major trauma, and other concomitant neoplasms), to ensure that the accelerated trajectories observed are genuinely associated with the tumour burden; (3) inconsistencies in diagnosis dates or cases in which CRC diagnosis could not be reliably confirmed; (4) appendiceal neoplasms; and (5) cases with multiple simultaneous metastatic sites or recurrent metastases within the same organ; and (6) patients exhibiting an extreme frequency of testing (>80 total tests) to prevent artificial leveraging of the longitudinal mixed models. Furthermore, regarding extreme outlier values (> ±3 standard deviations), these were individually and manually reviewed prior to exclusion. This was crucial to prevent geometric distortion in the mathematical models driven by isolated, acute physiological spikes rather than true longitudinal disease drift.

Haematological parameters

Data were extracted from the hospital’s central laboratory information system, consolidating results from both primary care and hospital settings. The dataset included red series parameters (Haemoglobin [Hb], RBC count, MCV, MCH, MCHC, RDW), white series (total leucocytes, neutrophils, lymphocytes, monocytes, eosinophils, basophils), platelet count and mean platelet volume (MPV), and derived indices such as NLR and PLR. Parameters were analysed across four retrospective intervals prior to diagnosis: T1 (0–1 month), T2 (2–4 months), T3 (5–12 months), and T4 (13–24 months). Where multiple tests existed per time window for a given patient, the arithmetic mean value was calculated to represent the period as a single representative value.

Statistical analysis

Data were analysed using R software (v4.1.2/4.4.0). Longitudinal trends were modelled using Linear Mixed-Effects Models (LMM) via the lme4 and lmerTest packages. The models were architected with fixed effects for the time window (T4 to T1), clinical group (G.0, G.Ia, G.Ib), and their interaction (Time × Group). To rigorously account for intra-individual variability and the correlated nature of repeated measures, a patient-specific random intercept was incorporated into the mathematical formulation (Value ∼ Time × Group + (1 | Patient_ID)). Model parameters and variance components were estimated using Restricted Maximum Likelihood (REML), and statistical significance (p-values) for the fixed effects was computed utilizing Satterthwaite's approximation for degrees of freedom.

Additionally, a Random Forest (RF) supervised learning model [40] was employed to evaluate variable importance and the discriminative power of haematological markers in distinguishing metastatic from non-metastatic phenotypes. Prior to model training, dynamic feature engineering was performed to explicitly capture the trajectory kinetics: relative temporal changes were calculated for erythroid parameters, while absolute differences were computed to map the velocity of change across leucocyte subsets, platelets, and derived inflammatory ratios. The refined dataset was split into an 80% training matrix and an independent 20% validation testing set. Model hyperparameter tuning was conducted via an automated grid-search using a 3-times repeated 5-fold cross-validation strategy to optimize internal calibration and prevent overfitting. The final predictive architecture consisted of 2000 bootstrap classification trees, a minimum leaf node size of 10, and an optimized classification probability threshold (cutoff = 0.84) tuned to maximise the Negative Predictive Value (NPV) required for clinical risk pre-staging.

Results

Longitudinal comparison between non-metastatic CRC (G.0) and synchronous/early hepatic metastases (G.Ia)

The Linear Mixed-Effects (LME) analysis identified consistent longitudinal signals distinguishing metastatic from non-metastatic phenotypes within the 24 months preceding diagnosis (Table 2). Crucially, the key discriminative information did not lie in absolute cross-sectional values, which frequently remained within normal reference ranges, but rather in the direction and magnitude of the underlying trajectories (Fig. 2). Statistical evaluation of the “Time × Group” interactions confirmed that patients developing synchronous liver metastases (G.Ia) exhibited significantly steeper temporal slopes across multiple parameters compared to the non-metastatic group (G.0), reflecting a progressive biological divergence.

Table 2.

Linear mixed-effects model results.

Parameter	Comparison	T4	T3	T2	T1
		p-value	p-value	p-value	p-value
Hb	G.0 vs G.Ia	<0.001	<0.001	<0.001	<0.001
	G.0 vs G.Ib	<0.001	<0.001	0.003	<0.001
RBC	G.0 vs G.Ia	<0.001	<0.001	<0.001	<0.001
	G.0 vs G.Ib	0.003	<0.001	0.005	<0.001
MCV	G.0 vs G.Ia	—	—	—	0.030
	G.0 vs G.Ib	0.01	0.03	—	0.006
MCH	G.0 vs G.Ia	0.06	0.02	—	0.02
	G.0 vs G.Ib	0.03	0.04	—	—
MCHC	G.0 vs G.Ia	—	0.04	0.03	—
	G.0 vs G.Ib	—	—	—	—
RDW	G.0 vs G.Ia	—	—	—	—
	G.0 vs G.Ib	—	—	0.02	—
Leucocytes	G.0 vs G.Ia	<0.001	<0.001	0.006	<0.001
	G.0 vs G.Ib	<0.001	<0.001	0.008	<0.001
Neutrophils	G.0 vs G.Ia	<0.001	<0.001	0.004	<0.001
	G.0 vs G.Ib	<0.001	<0.001	0.004	<0.001
Lymphocytes	G.0 vs G.Ia	<0.001	<0.001	0.007	<0.001
	G.0 vs G.Ib	<0.001	<0.001	<0.001	<0.001
Monocytes	G.0 vs G.Ia	—	—	—	—
	G.0 vs G.Ib	—	—	—	—
Eosinophils	G.0 vs G.Ia	—	—	—	—
	G.0 vs G.Ib	0.03	0.004	—	0.03
Basophils	G.0 vs G.Ia	—	—	0.02	0.007
	G.0 vs G.Ib	—	—	—	<0.001
Platelets	G.0 vs G.Ia	<0.001	<0.001	0.001	<0.001
	G.0 vs G.Ib	<0.001	<0.001	0.02	<0.001
MPV	G.0 vs G.Ia	—	—	—	—
	G.0 vs G.Ib	<0.001	0.006	—	<0.001
NLR	G.0 vs G.Ia	<0.001	<0.001	0.002	<0.001
	G.0 vs G.Ib	<0.001	<0.001	—	<0.001
PLR	G.0 vs G.Ia	<0.001	<0.001	<0.001	<0.001
	G.0 vs G.Ib	<0.001	<0.001	—	<0.001

Note:P-values obtained from longitudinal pairwise comparisons between CRC patients without metastases (G.0) and those with synchronous hepatic (G.Ia) or extrahepatic metastases (G.Ib) across predefined time intervals from T4 to T1 prior to CRC diagnosis. All comparisons were adjusted for multiple testing. Empty cells (—) indicate non-significant differences (p ≥ 0.05). Hb, haemoglobin; RBC, red blood cell count; MCV, mean corpuscular volume; MCH, mean corpuscular haemoglobin; MCHC, mean corpuscular haemoglobin concentration; RDW, red cell distribution width; MPV, mean platelet volume; NLR, neutrophil-to-lymphocyte ratio; PLR, platelet-to-lymphocyte ratio.

Fig. 2.

Longitudinal comparison of haematological trajectories across non-metastatic CRC (G.0), hepatic metastases (G.Ia), and extrahepatic metastases (G.Ib). Mixed-model estimated mean trajectories (±95% CI) for (a) haemoglobin (Hb), (b) red blood cells (RBC), (c) leucocytes, (d) neutrophils, (e) lymphocytes, (f) platelets, (g) neutrophil-to-lymphocyte ratio (NLR), and (h) platelet-to-lymphocyte ratio (PLR) across the four pre-diagnostic time periods (T4–T1). While absolute values largely overlap across groups, metastatic phenotypes exhibit steeper longitudinal deviations, defining a multidimensional metastatic gradient characterised by accelerated erythroid decline and progressive inflammatory activation prior to diagnosis. Note: NLR and PLR values are expressed on a logarithmic scale.

Regarding specific lineages, the erythroid series (haemoglobin and RBCs) showed a progressive decline towards diagnosis. This deterioration was significantly accelerated in the metastatic group, consistent with a pattern of emerging, tumour-burden-associated anaemia (Haemoglobin T4 interaction: β = 0.665, SE = 0.139, p < 0.001; RBC T4 interaction: β = 0.173, SE = 0.036, p < 0.001) (Table 3 and Fig. 3). Erythrocyte indices (MCV, MCH, MCHC, RDW) demonstrated temporal effects, although the interaction terms were less robust than in primary parameters (Table 3).

Table 3.

Longitudinal mixed-effects model estimates of haematological trajectory differences across metastatic phenotypes.

Parameter	Trend (T4→T1)	β ± SE (T4xG.Ia)	p(Ia)	β ± SE (T4xG.Ib)	p(Ib)	Clinical Interpretation
Haemoglobin (Hb)	↓	0.665 ± 0.139	<0.001	0.521 ± 0.139	<0.001	Steeper Hb declines (anaemia trajectory) in metastatic groups.
Erythrocytes (RBC)	↓	0.173 ± 0.036	<0.001	0.109 ± 0.035	0.002	Greater decrease in RBC in metastatic groups.
MCV	↓	—	—	1.245 ± 0.488	0.011	Slight decrease, signal clearer in G.Ib.
MCH	↓	—	—	0.458 ± 0.209	0.028	Slight decrease, signal clearer in G.Ib.
MCHC	↓	—	—	—	—	Minimal/Nonsignificant interaction.
RDW	↑	—	—	—	—	Increasing anisocytosis toward diagnosis. Nonsignificant interaction.
Leukocytes	↑	−1.182 ± 0.279	<0.001	−1.439 ± 0.278	<0.001	Steeper leukocytes increase in metastatic groups. Systemic inflammatory activation.
Neutrophils	↑	−4.310 ± 0.821	<0.001	−4.589 ± 0.819	<0.001	Steeper neutrophil increase in metastatic groups.
Lymphocytes	↓	4.175 ± 0.680	<0.001	4.061 ± 0.678	<0.001	Steeper lymphocyte declines toward diagnosis in metastatic groups. Progressive lymphopenia.
Monocytes	↓	—	—	—	—	Nonsignificant interaction.
Eosinophils	↓	—	—	0.323 ± 0.146	0.027	Mild eosinophil reduction.
Basophils	↓	—	—	—	—	Nonsignificant interaction.
Platelets	↑	−42.275 ± 6.164	<0.001	−27.336 ± 6.142	<0.001	Steeper platelet rise in metastatic groups, thrombocytosis.
MPV	↓	—	—	0.250 ± 0.090	0.005	Moderate signal, stronger in G.Ib.
NLR	↑	−1.512 ± 0.296	<0.001	−1.032 ± 0.296	<0.001	Stronger NLR rise in metastatic groups.
PLR	↑	−0.652 ± 0.078	<0.001	−0.447 ± 0.078	<0.001	Stronger PLR rise in metastatic groups.

Note: β, fixed-effect interaction coefficient estimates from REML models (Satterthwaite approximation); SE, Standard Error. All models include a random intercept per patient (1|NHC). Time baseline is T1; coefficients represent differential slopes at the final pre-diagnostic window (T4) for Groups Ia (G.Ia) and Ib (G.Ib) relative to controls. NLR and PLR were log-transformed (coefficients expressed in logarithmic units). Hb, haemoglobin; RBC, red blood cell count; MCV, mean corpuscular volume; MCH, mean corpuscular haemoglobin; MCHC, mean corpuscular haemoglobin concentration; RDW, red cell distribution width; MPV, mean platelet volume. Em-dashes (—) indicate non-significant interaction effects (p ≥ 0.05).

Fig. 3.

Forest plot of temporal slope estimates ($\mathit{\beta }$) for significant haematological parameters across metastatic phenotypes. Fixed-effect interaction coefficients ($\mathit{\beta }$) and 95% confidence intervals (CI) derived from the Linear Mixed-Effects models. The dashed vertical line at zero represents the baseline temporal trajectory of the non-metastatic reference group (G.0). Point estimates indicate the differential slope magnitude for synchronous hepatic (G.Ia) and extrahepatic (G.Ib) metastases during the pre-diagnostic window. Due to the retrospective temporal coding (measured backwards from the time of diagnosis), slope directionality is inverted: positive $\mathit{\beta }$ values denote a progressive decline in parameter levels as the diagnostic event approaches (historical values > recent values), whereas negative $\mathit{\beta }$ values indicate a progressive increase towards diagnosis (historical values < recent values). The plot is stratified into primary erythrocyte and leucocyte lineages (top), platelet count (middle), and derived inflammatory ratios (bottom). Note: NLR and PLR coefficients are expressed in logarithmic units. Hb, haemoglobin; RBC, red blood cells; Leu, leucocytes; Neu, neutrophils; Lin, lymphocytes; Plq, platelets; NLR, neutrophil-to-lymphocyte ratio; PLR, platelet-to-lymphocyte ratio; CI, confidence interval.

Conversely, the white cell series displayed a divergent pattern driven by systemic inflammatory activation. Total leucocytes and neutrophils increased, with this expansion being significantly more pronounced in the hepatic metastatic cases compared to the non-metastatic baseline (Leucocytes: β = −1.182, SE = 0.279, p < 0.001; Neutrophils: β = −4.310, SE = 0.821, p < 0.001). Simultaneously, lymphocytes exhibited a marked downward trajectory, which was significantly steeper in the G.Ia cohort (β = 4.175, SE = 0.680, p < 0.001). Similarly, platelets showed a gradual increase, with a significantly sharper thrombocytosis trajectory observed in the G.Ia group (β = −42.275, SE = 6.164, p < 0.001) (Table 3 and Fig. 3).

These trends culminated in the derived indices, NLR and PLR, which demonstrate the strongest “Time × Group” interactions. These indices revealed a highly accelerated expansion of the inflammatory and pro-metastatic response, alongside lymphocytic exhaustion, in patients with liver metastases, exhibiting pronounced slope divergences at T4 (NLR: β = −1.512, SE = 0.296, p < 0.001; PLR: β = −0.652, SE = 0.078, p < 0.001) (Table 3 and Fig. 3).

Expanded analysis including extrahepatic synchronous/early metastases (G.Ib)

To assess whether the observed trends extended beyond hepatic disease, the analysis incorporated patients with synchronous/early extrahepatic metastases (G.Ib). Significant “Time × Group” interactions persisted, confirming that both metastatic cohorts consistently diverged from the non-metastatic trajectory (G.0) (Table 2 and Fig. 2). However, the specific haematological profile varied by metastatic site, suggesting subtle differences in the systemic response.

Leucocytes and neutrophils exhibited the most pronounced upward trajectories in the extrahepatic metastases group (G.Ib), outpacing even the hepatic metastases group (G.Ia) (Leucocytes: β = −1.439, SE = 0.278, p < 0.001; Neutrophils: β = −4.589, SE = 0.819, p < 0.001) (Table 3 and Fig. 3). This pattern clearly separates both metastatic phenotypes from the non-metastatic baseline (G.0).

Conversely, the marked thrombocytosis observed in G.Ib (β = −27.336, SE = 6.142, p < 0.001) suggests that extrahepatic dissemination may be partially driven by platelet-mediated protection and aggregation mechanisms, yet the interaction slopes confirm this platelet rise was significantly more pronounced in the hepatic cohort (G.Ia). Similarly, the inflammatory response mediated by the NLR index (β = −1.032, SE = 0.296, p < 0.001) was notably intense in the extrahepatic group, but also displayed a steeper trajectory in the hepatic cohort (G.Ia) (Table 3 and Fig. 3).

Regarding the PLR, both metastatic cohorts displayed a parallel and significant escalation relative to G.0 (G.Ib PLR interaction: β = −0.447, SE = 0.078, p < 0.001); this phenotypic similarity in the derived index, however, reflects a distinct dynamic equilibrium within each group. While the magnitude of the ratio in the hepatic group (G.Ia) is heavily magnified by the combination of critical thrombocytosis and severe lymphopenia, the rise in extrahepatic cases (G.Ib) maintains an equivalent clinical relevance driven by sustained peripheral inflammatory stimulation, despite exhibiting slightly more attenuated platelet proliferation curves than those observed in the hepatic cohort (Table 3 and Fig. 2, Fig. 3).

Finally, the “loss of function” signals, specifically lymphocytes and erythroid parameters, showed progressive declines across both metastatic cohorts. This deterioration in the red cell series is consistent with iron sequestration mechanisms characteristic of chronic systemic inflammation, which impairs erythropoiesis. Notably, the lymphopenic trajectory was steeper in the hepatic group (G.Ia) compared to the extrahepatic group (G.Ib: β = 4.061, SE = 0.678, p < 0.001), reinforcing the concept of a profound immune exhaustion specifically associated with the liver metastatic microenvironment.

These findings support the presence of a multidimensional metastatic gradient. Table 3 provides an integrated synthesis of these longitudinal effects, detailing the direction of trends, effect sizes ($\beta$), and clinical interpretation for each marker.

Machine-learning classification models (Random forest)

To complement the longitudinal mixed-model analyses and evaluate the discriminative potential of these trajectories, Random Forest (RF) classifiers were developed. The model comparing non-metastatic patients (G.0) against the combined metastatic group (G.Ia–G.Ib) yielded the most balanced performance, achieving an AUC of 0.74, a sensitivity of 73.0%, and a specificity of 61.0% (Table 4).

Table 4.

Random Forest performance metrics for comparisons between non-metastatic CRC (G.0) and metastatic groups (G.Ia, G.Ia–G.Ib).

	AUC	Sensitivity	Specificity	PPV	NPV	Balanced accuracy
RF G.0 vs G.Ia	0.73–0.74	67.3%	65.6%	27.1%	91.4%	64.4%
RF G.0 vs G.Ia-Ib	0.73–0.74	73.0%	61.0%	38.6%	87.0%	67.0%

Clinical Implication of Prediction. While the discriminative ability is statistically moderate—which is consistent with the challenge of classifying early evolutionary stages of the same underlying disease—the model achieved a robust Negative Predictive Value (NPV) of 87.0%. This metric is clinically invaluable for pre-staging, as it highlights a strong capacity to confidently exclude low-risk patients. Although the Positive Predictive Value (PPV) is modest (38.6%) due to the naturally low prevalence of metastasis, this performance remains sufficient to flag a high-risk subgroup deserving intensified radiological surveillance.

Model Stability and Robustness. The reliability of these predictions was further corroborated by the stability analysis shown in Fig. 4. The forest plot demonstrates narrow confidence intervals for the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity estimates across repeated calibration folds. This clustering of results confirms that the classifier is robust, demonstrating that the predictive capacity is a consistent signal rather than an artifact of specific data partitioning.

Fig. 4.

Random Forest hyperparameter tuning and classification performance. Performance metrics (Area Under the Receiver Operating Characteristic curve [ROC], Sensitivity [Sens], and Specificity [Spec]) for various Random Forest configuration sets, comparing non-metastatic (G.0) versus metastatic (G.Ia–G.Ib) phenotypes. Each row represents a specific hyperparameter combination (mtry, nodesize). Data are presented as point estimates with 95% confidence intervals. Confidence level: 0.95.

Variable Importance and Validation. Variable-importance analyses (Fig. 5) revealed a coherent pattern: lymphocyte counts (particularly at T1), neutrophils, NLR, and PLR emerged as the most influential predictors. This ranking directly mirrors the findings of the mixed-effects model, reinforcing the dominant role of lymphopenia, myeloid activation, and platelet-driven inflammation as core components of the metastatic haematological profile. The concordance between the machine-learning outputs and the longitudinal analyses strongly supports the biological validity of this pre-diagnostic signature.

Fig. 5.

Variable importance rankings in the Random Forest model (G.0 vs. G.Ia–G.Ib). Predictor importance is determined by MeanDecreaseAccuracy (left) and MeanDecreaseGini (right). Feature labels denote specific time periods or longitudinal changes: suffix numbers indicate the time window (e.g., 1 = T1, 2 = T2, 3 = T3, 4 = T4), while dual-number suffixes represent the longitudinal difference between time points (e.g., Lin12 = difference between T1 and T2). Higher values indicate greater contribution to classification performance. Lymphocyte and neutrophil measures, together with derived inflammatory indices (NLR and PLR), ranked among the most influential predictors, underscoring the dominant role of longitudinal immune and inflammatory dynamics in metastatic risk discrimination. Lin, lymphocytes; Neu, neutrophils; Leu, leucocytes; Plq, platelets; Eos, eosinophils; Mon, monocytes; Hb, haemoglobin; MCV, mean corpuscular volume; RDW, red cell distribution width; NLR, neutrophil-to-lymphocyte ratio; PLR, platelet-to-lymphocyte ratio; MPV, mean platelet volume; AgeDiag, patient age at diagnosis.

Discussion

Building upon our prior work on the red blood cell series [41], this study broadens the analysis to the complete blood count (CBC), revealing that CRC is preceded by a systemic haematological signature detectable at least 24 months before diagnosis. By implementing linear mixed-effects models (LMM), we identified a "metastatic gradient": a progressive divergence where patients presenting with synchronous metastases (G.Ia and G.Ib) exhibit significantly accelerated trajectories, haemoglobin decline and inflammatory expansion, compared to non-metastatic cases (G.0).

Biological Plausibility of the Longitudinal Trajectories. The observed deterioration in haemoglobin supports previously described longitudinal trends preceding CRC diagnosis [42,43]. Importantly, Li et al. reported a non-linear acceleration of haemoglobin decline as diagnosis approaches and suggested that such temporal patterns, rather than single measurements, may reflect underlying disease progression. In line with this hypothesis, our data indicate that anaemia is not merely a consequence of blood loss, but a reflection of tumour burden and systemic progression, as evidenced by its steeper longitudinal decline in metastatic patients [44]. Concomitantly, the immune shift, characterised by neutrophilia and lymphopenia, reflects the host’s transition from immunosurveillance to a pro-tumorigenic environment. The neutrophil expansion aligns with their role in promoting metastasis through neutrophil extracellular traps (NETs) [26,27], while the progressive thrombocytosis supports the concept of “tumour-educated platelets” (TEPs), which facilitate immune evasion and epithelial–mesenchymal transition [45,46].

These variations are magnified by the derived indices NLR and PLR. Our results confirm that these indices can be robust prognostic markers [47,48], but with a crucial temporal distinction: the prognostic divergence does not begin at diagnosis. Instead, the accelerated escalation of NLR and PLR establishes itself progressively months in advance. These findings align with the hypothesis by Massagué and Obenauf, who propose that metastatic dissemination may commence during a latent phase, long before the primary tumour is clinically apparent [49]. Importantly, these trends often unfold within "normal" reference ranges, therefore assigning to CBC a possible role as a "biological chronometer" in which trajectory is more informative than absolute value.

Clinical Utility and Scope. The Random Forest model (AUC 0.74) operates as a haematological pre-staging tool for risk stratification. Its robust Negative Predictive Value (NPV, 87.0%) proves particularly useful in ruling out high-risk phenotypes, supporting clinical decision-making in patients whose standard blood counts remain within normal ranges but whose longitudinal trajectories indicate biological progression. The observed Positive Predictive Value (PPV, 38.6%) must be interpreted within the biological context of the study: the model differentiates between two evolutionary stages of the same underlying neoplastic disease rather than distinguishing healthy individuals from cancer patients. Because the primary colorectal tumour independently drives systemic inflammation and anaemia, a high degree of biological overlap is expected, which inherently caps predictive precision. Consequently, this tool effectively flags a high-risk subgroup warranting intensified, risk-adapted radiological staging, complementing established diagnostic procedures.

Limitations and Future Directions. This study presents specific limitations inherent to its retrospective, single-centre design, which warrants caution in generalising findings to highly heterogeneous populations. While documented cases of severe alternative anaemia or inflammatory comorbidities were rigorously excluded, unmeasured non-oncologic confounders remain a potential source of variance. Furthermore, the focus on CRC metastatic phenotypes necessitates broader investigation to determine the definitive real-world scope of these haematological trajectories. Future studies evaluating other neoplasms and chronic inflammatory conditions are essential to confirm the disease-specificity of these patterns. It is within this broader context that the true utility of the model will be refined; the future inclusion of healthy control cohorts and groups with divergent biological behaviour will likely enhance the discriminative capacity of these longitudinal algorithms. Finally, as this predictive model lacks external validation and Decision Curve Analysis (DCA) is not feasible in a retrospective framework, these findings serve as a proof-of-concept. The present model should not be interpreted as ready for clinical use. Its performance metrics, although internally stable, must be recalibrated and tested in prospective, multicentre studies before any integration into diagnostic pathways is considered.

Conclusions

Our findings substantiate the clinical utility of longitudinal blood count monitoring as a non-invasive tool for haematological pre-staging in colorectal cancer. Rather than relying on static reference ranges, tracing patient-specific dynamic trajectories allows for the early identification of accelerated inflammatory and thrombotic signals that anticipate aggressive synchronous metastatic phenotypes. The current machine-learning model demonstrates robust internal stability and a high negative predictive value, providing a foundation for future clinical implementation. However, while our results highlight the potential of longitudinal CBC trajectories, the approach remains investigational and cannot be applied clinically until prospective external validation confirms its reproducibility and clinical benefit.

Ethics approval and consent to participate

This study adhered to the principles of the Declaration of Helsinki and was approved by the Ethics and Clinical Research Committee of La Ribera University Hospital (26 February 2018; registry code HULR20181026). All patient data were pseudonymised in strict compliance with the General Data Protection Regulation (GDPR, EU 2016/679).

Given the retrospective nature of the study and the use of anonymised data extracted from hospital medical records under the supervision of quality management personnel, retrospective informed consent was deemed unnecessary and this was approved by the Ethics Committee.

Generative AI and AI-assisted technologies statement

During the preparation of this work, the authors used AI-assisted tools including Scopus AI, SciSpace, and OpenEvidence to support literature searching and content organisation. In addition, ChatGPT (OpenAI) and Gemini (Google) were used exclusively for language translation, grammatical correction, and stylistic refinement during manuscript preparation. Following the use of these tools, the authors critically reviewed and revised all material as necessary and accept full responsibility for the accuracy, integrity, and content of the published work.

Funding

This research received no external funding.

CRediT authorship contribution statement

Rafael J. Sala: Writing – review & editing, Writing – original draft, Resources, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. John Ery: Methodology, Investigation, Formal analysis, Data curation. David Cuesta-Peredo: Resources, Data curation. Vicente Muedra: Writing – review & editing, Writing – original draft, Supervision, Funding acquisition, Formal analysis, Conceptualization. Vicent Rodilla: Writing – review & editing, Writing – original draft, Supervision, Methodology, Formal analysis, Conceptualization.

Declaration of competing interest

The authors declare that they have no competing interests.

Data availability

Data were obtained from hospital medical records and may be made available upon reasonable request to the authors, subject to approval by the hospital authorities.