AI Diagnoses Multiple Diseases From a Single Blood Sample

Medical diagnosis can sometimes be complex. When different diseases have similar symptoms, multiple laboratory tests may be needed to reach the correct diagnosis, making the process time-consuming and expensive. Recent advances in artificial intelligence (AI) offer innovative approaches that could address this challenge.

A study published on February 21, 2025, in Science developed a remarkable AI tool capable of diagnosing multiple diseases from a single blood sample.

Decoding the Immune System's Hidden Language

Stanford University researchers developed an algorithm that analyzes genes encoding immune cell receptors to accurately identify diseases such as COVID-19, HIV, type 1 diabetes, and lupus, as well as responses to influenza vaccination.

These diseases and conditions represent different aspects of the immune system: viral infections, autoimmune diseases, post-vaccination responses, and healthy immune status. They can be difficult to distinguish or produce similar immune responses. They were selected to comprehensively test the AI model's ability to detect subtle differences in immune responses.

Called “Machine Learning for Immunological Diagnosis” (Mal-ID), the algorithm diagnoses conditions by analyzing genetic data from B-cell and T-cell receptors.

One of the study's senior authors, Prof. Scott Boyd, emphasized its importance: “Immune cell receptor sequences provide unique diagnostic information. Until now, these sequences have generally only been used to diagnose cancers originating from B or T cells.”

The illustration above summarizes step by step how the new AI tool analyzes blood samples to diagnose diseases. Here is a detailed explanation:

Step 1: Obtaining B-Cell and T-Cell Receptor Sequences

• The researchers determined B-cell (BHR) and T-cell (THR) receptor sequences from blood samples collected from 593 participants.
• Participants represented six immune states: COVID-19, HIV, lupus, type 1 diabetes, an immune response after influenza vaccination, and healthy individuals.

2. Training Three Separate AI Models

• Receptor sequences were examined using three approaches:

  1. Repertoire Composition: The overall distribution of B-cell and T-cell receptors in each person's immune system was analyzed.
  2. CDR3 Sequence Clustering: Critical regions determining antigen recognition (CDR3) were examined using clustering.
  3. Language Modeling: Language models analyzed receptor sequences to decipher the immune system's linguistic features.

• Data from the three approaches were used to train three BHR and three THR models.

Creating a Combined Machine Learning Model

• The trained models were combined into a comprehensive machine learning (ML) model.

Diagnosing Diseases and Interpreting Results

• The machine learning model was evaluated through the following steps:
  • Cross-Validation: The model's accuracy was assessed.
  • Testing With Independent Datasets: Performance was validated using external datasets.
  • Comparison With the Literature: The most informative BHR and THR sequences were compared with published information to validate the findings.

This approach enabled highly accurate diagnosis, with an AUROC score of 0.986, and provided a comprehensive picture of the immune system. In the future, it may allow many diseases to be diagnosed rapidly and accurately with a single test.

Key Findings

The study analyzed receptor sequences from a total of 593 people, comprising 23.5 million T-cell and 16.2 million B-cell receptor sequences. Participants included:

• 63 patients with COVID-19

• 95 patients with HIV

• 86 patients with lupus

• 92 patients with type 1 diabetes

• 37 people who had recently received an influenza vaccine

• 220 healthy controls.

The algorithm's performance was notable. Using both B-cell and T-cell data, it achieved 85.3% accuracy and an AUROC (area under the receiver operating characteristic curve) of 0.986. With B-cell data alone, accuracy was 74.0% and AUROC was 0.959; with T-cell data alone, accuracy was 75.1% and AUROC was 0.952.

These results showed that combining B-cell and T-cell data improved diagnostic sensitivity. For example, B-cell receptors were more informative for COVID-19, HIV, and influenza vaccination, whereas T-cell receptors provided more distinctive signals in lupus and type 1 diabetes.

AI and the Immune System: Discovering Dialects

Immune system receptors resemble unique dialects that record each person's immune history. AI language models, such as the large language models behind ChatGPT, are particularly effective at understanding these distinctive dialects.

Potential Clinical Applications and Future Directions

Although this innovative approach is not yet ready for clinical use, the researchers believe it could substantially reduce the time and cost of diagnosis. Analyzing immune receptor sequences could also lay the groundwork for personalized treatment approaches to autoimmune diseases.

Study author Maxim Zaslavsky notes that AI has considerable potential to identify the most effective treatments for complex conditions such as autoimmune diseases by analyzing immune activity.

The study is regarded as an important step that could transform diagnosis and treatment through detailed analysis of the immune system.