Deep phenotyping of blood cell data reveals novel clinical biomarkers

This study demonstrates that applying AI techniques, specifically clustering and self-supervised autoencoders, to raw single-cell blood count data can uncover novel, clinically prognostic biomarkers that capture subtle physiologic signals and predict outcomes like mortality and disease development more effectively than traditional summary metrics.

The Gist

Imagine your blood is a bustling city, and the Complete Blood Count (CBC) test is the daily newspaper that the hospital reads to check on the city's health.

For decades, this newspaper has only reported the "headlines": How many people are there? What is the average size of a house? How many police officers are on patrol? These are the standard numbers doctors look at (like total white blood cell count or average red blood cell size).

The Problem:
The authors of this paper realized that while the "newspaper" gives us the headlines, the actual raw data collected by the machines is like a high-definition, 4K video recording of every single person walking down the street. The machine sees thousands of tiny details about every single cell—how bumpy it is, how much light it reflects, how it's moving. But until now, doctors have been throwing away this rich video footage and only reading the summary headlines. They were missing subtle shifts in the crowd that could signal trouble before a crisis happens.

The Solution: Two New Ways to Watch the Video
The researchers, led by Dr. Brody Foy, decided to use Artificial Intelligence (AI) to re-watch this raw video footage and find new, hidden stories. They used two different "lenses" to do this:

1. The "Organized Crowd" Lens (Clustering):
   Imagine sorting the crowd in the video into specific groups: the firefighters, the teachers, the construction workers. Once sorted, instead of just counting them, the AI looks at the details of each group.

   • The Analogy: Instead of just saying "There are 100 firefighters," the AI says, "The firefighters are unusually jittery," or "The smallest firefighters are getting much smaller."
   • The Result: They found that looking at the variance (how different the cells are from each other) and the extremes (the very smallest or largest cells) was a powerful predictor of future illness. It's like noticing that the smallest monocytes (a type of white blood cell) are shrinking, which predicts heart trouble years before a heart attack.

2. The "Pattern Detective" Lens (Autoencoders):
   This is the more mysterious AI. Imagine a detective who doesn't care about specific groups but looks for complex, invisible patterns connecting everyone in the crowd. Maybe the way the construction workers are standing relates to how the police are moving, even though they aren't talking to each other.

   • The Analogy: This AI finds "secret codes" in the data that humans can't easily explain. It captures non-linear relationships—like a complex dance between different cell types that signals stress in the body.
   • The Result: These "secret codes" turned out to be incredibly good at predicting who would get sick, who would need to be admitted to the hospital, or who might develop cancer, often better than the standard headlines.

What Did They Discover?
By applying these AI lenses to over 240,000 blood tests, they found:

• New Early Warning Systems: They discovered hundreds of new "biomarkers" (warning signs). For example, they found that the spread of sizes in neutrophils (a type of white blood cell) is a strong predictor of death or hospital admission.
• Hidden Connections: The "Pattern Detective" AI found that these blood cell patterns were secretly linked to other things in the body, like specific infections (like HIV or CMV), hormone levels, and even how well your blood clots. It's like realizing that the way the "firefighters" are walking in your blood is actually a sign that your "plumbing" (coagulation) is stressed.
• Better Than the Headlines: Even after adjusting for age, sex, and the standard blood test numbers, these new AI-derived markers still provided strong predictions. They added a whole new layer of information that was previously invisible.

Why Does This Matter?
Think of it like upgrading from a black-and-white radio to a surround-sound 3D movie.

• Current Medicine: We listen to the radio (standard blood counts). It tells us if there's a storm, but only after the wind starts blowing.
• This Study: We are now watching the 3D movie. We can see the clouds gathering and the pressure dropping before the storm hits.

The Bottom Line:
This paper proves that we have been sitting on a goldmine of data for years. By using modern AI to dig deeper into the raw, single-cell data from routine blood tests, we can create new, highly accurate tools to predict disease, catch problems earlier, and potentially save lives—all without needing to invent new, expensive tests. We just needed better software to read the data we already have.

Technical Summary

1. Problem Statement

The Complete Blood Count with Differential (CBD) is a ubiquitous clinical test worldwide. While modern analyzers utilize flow cytometry to generate rich, single-cell data (tens of thousands of cells per sample), clinical reporting is restricted to coarse summary statistics (e.g., total cell counts, mean cell size, and average fluorescence).

• The Gap: This aggregation discards subtle shifts in cell population distributions (e.g., variance, tail behaviors, and non-linear interactions) that may signal early-stage pathogenesis.
• The Limitation of Current Solutions: Existing "research parameters" (Cell Population Data or CPD) are often proprietary, lack technical transparency, and are not systematically evaluated. Prior attempts to analyze raw single-cell data have relied on manual gating or focused on specific diseases rather than systematic, unbiased biomarker discovery.

2. Methodology

The authors developed a two-pronged approach to extract novel biomarkers from raw single-cell flow cytometry data (.fcs files) collected from the University of Washington Medical Center (Sysmex XN-1000 analyzers).

A. Data Source & Cohort

• Dataset: 242,623 CBD samples from 127,545 patients (Apr 2024 – Aug 2025).
• Channels: Four analyzer channels were analyzed: WNR (white cells/basophils), WDF (white cell differential), PLTC (platelets), and RETC (red cells/reticulocytes).
• Features: Each cell was characterized by four scatter features: Forward Scatter (FSC), Side Scatter (SSC), Forward Scatter Width (FSC-W), and Side Fluorescence (SFL).
• Split: Data was split into a training set (first year) and a test set (last 5 months).

B. Approach 1: Interpretable Biomarkers (CLS)

• Clustering: Used FlowSOM (Self-Organizing Maps followed by consensus hierarchical clustering) to identify cell sub-populations without manual gating.
• Feature Extraction: For each identified cluster (e.g., neutrophils, lymphocytes), robust statistical summaries were computed across 9 percentiles (1st–99th), mean, standard deviation, min/max, and pairwise covariances of the four scatter features.
• Output: ~58 features per cell type, resulting in hundreds of interpretable markers (e.g., "variance of neutrophil side fluorescence").

C. Approach 2: Non-Linear Biomarkers (AE)

• Modeling: Trained three types of Self-Supervised Autoencoders to learn compact, non-linear representations of the raw cell data:
  1. Feedforward (FF): Fully connected layers.
  2. Convolutional (CNN): Convolutional layers with pooling.
  3. Set Model (Transformer-based): Implemented the Fully Differentiable Set Autoencoder (FDSA) framework. This model treats cells as unordered sets and uses attention mechanisms to explicitly model cell-cell interactions.
• Training: Models were trained to reconstruct input data, with the Set model achieving the lowest reconstruction error (MSE).
• Output: Compact embedding vectors representing complex, non-linear physiological states.

D. Evaluation Strategy

• Outcomes: Associations were tested against five major outcomes: Inpatient (IP) admission, 30-day mortality, and future diagnosis of Iron Deficiency Anemia, Cancer, and Major Adverse Cardiovascular Events (MACE).
• Statistical Models: Logistic regression (for admission/mortality) and Cox proportional hazards (for disease diagnosis), adjusted for demographics and standard CBC markers.
• Novelty Assessment: Correlations were calculated between new markers and existing CBD/CPD markers. "Novel" markers were defined as having low correlation (<0.5) with existing markers while retaining significant prognostic value.

3. Key Results

• Clustering Accuracy: The unsupervised clustering successfully recapitulated known cell types, showing correlation coefficients >0.9 with analyzer-reported counts for most cell types (except reticulocytes, which showed continuous maturation).
• Prognostic Power:
  • CLS Markers: Hundreds of new markers showed significant associations with all outcomes. Notably, variance and tail distribution markers (e.g., 1st percentile of monocyte size) were stronger predictors than mean values. For example, neutrophil SFL variance predicted mortality (OR=1.53) and IP admission (OR=1.42).
  • AE Markers: The Set autoencoder (FDSA) outperformed FF and CNN models in reconstruction and identified distinct patient subgroups. It generated 169 novel clinical markers.
• Novelty & Independence:
  • Both CLS and AE markers remained significant after adjusting for standard CBC indices (RBC, WBC, PLT, RDW) and demographics.
  • A significant subset of markers had low correlation (<0.5) with existing CBD/CPD markers, proving they capture unique physiological signals.
• Physiologic Correlations:
  • AE markers correlated significantly with broader non-CBD laboratory tests, including specific immune cell subsets (CD4+), infectious disease markers (CMV, HIV), coagulation factors (Factor VIII), and anemia markers (G6PD, Zinc-protoporphyrin).
  • This suggests the AI models captured systemic signals (inflammation, coagulation, infection) embedded within routine blood counts.

4. Key Contributions

1. Systematic Deep Phenotyping: First large-scale study to systematically extract novel biomarkers from raw single-cell CBD data using both interpretable (clustering) and non-linear (deep learning) methods.
2. Methodological Framework: Established a pipeline combining FlowSOM for biological interpretability and Set-based Autoencoders (FDSA) for capturing complex cell-cell interactions.
3. Discovery of "Tail" Signals: Demonstrated that the most clinically relevant signals often reside in the tails of cell distributions (variance, extreme values) rather than mean values, which are already captured by standard tests.
4. Clinical Utility: Generated hundreds of new biomarkers that predict mortality, hospitalization, and specific diseases (cancer, MACE, anemia) independently of current standard-of-care tests.

5. Significance

• Enhanced Diagnostics: This approach unlocks the "dark matter" of routine blood tests, turning a standard, low-cost test into a source of high-dimensional, multi-system health data.
• Early Detection: By detecting subtle distributional shifts (variance/tails) before mean values change, these markers could enable earlier detection of pathogenesis (e.g., pre-anemic states or early cardiovascular risk).
• Interpretability vs. Performance: The study successfully bridges the gap between "black-box" AI and clinical utility. The CLS markers offer clear biological interpretation (e.g., "increased neutrophil size variance"), while AE markers provide superior predictive power for complex, non-linear conditions.
• Scalability: Since flow cytometry is standard in almost all clinical labs, this methodology can be deployed globally to improve risk stratification without requiring new hardware or expensive tests.

Conclusion: The paper demonstrates that applying AI to raw flow cytometry data from routine blood counts reveals novel, prognostic biomarkers that capture subtle physiologic signals missed by current clinical standards, offering a pathway to more precise and early disease detection.