Hierarchical Machine Learning for Reproducible Cytometry Cell Classification

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional cytometry techniques for identifying cell types are subjective, inefficient, and produce inconsistent results due to manual data analysis, leading to inaccuracies and challenges in reproducibility and data comparison across studies.

Innovation Solution

Utilizing multiple machine learning models, including a first model to classify events as cells, debris, or beads, and a second model to determine specific cell types, improving accuracy and reproducibility by eliminating human subjectivity and enhancing cell type differentiation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If manual data analysis is used in conventional cytometry techniques, then the process is simple to implement, but the results are subjective, inconsistent, and inefficient

Engineering Contradiction:
Improvecell type identification accuracyVSAvoiddata analysis system complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent replaces manual mechanical analysis with automated machine learning models. The first machine learning model automatically classifies events as cells, debris, or beads, and the second machine learning model determines specific cell types, eliminating human subjectivity and improving measurement precision while accepting increased system complexity through automated computational processing

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The machine learning models perform self-service by automatically processing cytometry data without human intervention. The models independently classify events and determine cell types based on training data, providing consistent and reproducible results that improve measurement precision while reducing reliance on manual analysis

Inventive Principle:
Principle #25Self-service

2Reliability

If multiple machine learning models are used to classify events and determine cell types, then measurement precision and reproducibility improve, but device complexity and computational requirements increase

Engineering Contradiction:
Improvedata reproducibilityVSAvoidmachine learning model complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent segments the data analysis process into two distinct machine learning models: the first model segments events into categories (cells, debris, beads), and the second model segments cell identification into specific cell types. This segmentation improves reliability by assigning specialized functions to each model, while the modular structure manages complexity through divided computational tasks

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The first machine learning model acts as an intermediary between raw cytometry data and the second machine learning model. It pre-processes data by classifying events into categories, filtering out non-cell events before they reach the second model. This intermediary step improves reliability by ensuring the second model only processes relevant cell data, while managing overall system complexity through staged processing

Inventive Principle:
Principle #24Intermediary (Mediator)

3Productivity

If manual gating and analysis are performed by operators, then the system is easier to operate, but productivity and efficiency are reduced

Engineering Contradiction:
Improvedata processing speedVSAvoidsystem operation simplicity
Core Design Contradiction:
ProductivityVSEase of operation

Solution Approach 1:

The patent replaces manual operator gating and analysis with automated machine learning models that process cytometry data at high speed. The models automatically classify events and determine cell types without human intervention, dramatically improving productivity and data processing speed while reducing the simplicity of operation through automated computational systems

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The machine learning models perform self-service by automatically processing cytometry data without requiring operator intervention for each analysis. The models independently execute classification and cell type determination, improving productivity through continuous automated processing while reducing operational simplicity through autonomous computational decision-making

Inventive Principle:
Principle #25Self-service

Data Source

PatentEP4473288B1Machine learning techniques for cytometry
Publication Date: 2025.10.29 BOSTONGENE CORP
  • EP4473288B1 patent drawingFigure 1A
  • EP4473288B1 patent drawingFigure 1B~1C
  • EP4473288B1 patent drawingFigure 1D

AI summary

Techniques for determining a respective cell type for each of at least some of a plurality of cells. The techniques includes: obtaining cytometry data for a biological sample from a subject, the biological sample comprising a plurality of cells including a first cell, the cytometry data including first cytometry data for the first cell; and determining a respective type for each of at least some of the plurality of cells using a hierarchy of machine learning models corresponding to a hierarchy of cell types, the determining comprising determining a first type for the first cell by processing the first cytometry data using a first subset of the hierarchy of machine learning models.