Short answer

When developing AI for diagnostic applications, prioritize explainability to ensure transparency and alignment with expert knowledge, thereby fostering user trust and facilitating clinical integration.

Field: Commercial Production
Source: PLOS Digital Health (2023)
Method: Machine Learning (Explainable AI, Multiple Instance Learning)
Sample: 189 patients (129 AML, 60 controls)
Evidence: Strong effect

An explainable AI model can accurately classify acute myeloid leukemia subtypes from blood smears by focusing on diagnostically relevant cells, mirroring human expert identification. This commercial production research insight is drawn from a 2023 study published in PLOS Digital Health. Using Machine learning (explainable ai, multiple instance learning) with 189 patients (129 AML, 60 controls), researchers explored how this design variable affects real-world outcomes. The key design takeaway: When developing AI for diagnostic applications, prioritize explainability to ensure transparency and alignment with expert knowledge, thereby fostering user trust and facilitating clinical integration.

Study

Commercial ProductionRecentStrong effect

AI-driven cell analysis achieves 94% accuracy in AML subtype diagnosis, matching expert pathologist focus

An explainable AI model can accurately classify acute myeloid leukemia subtypes from blood smears by focusing on diagnostically relevant cells, mirroring human expert identification.

PLOS Digital Health · 2023

Key Findings

01SCEMILA perfectly discriminated between AML patients and healthy controls.
02The model detected the APL subtype with an F1 score of 0.86 ± 0.05.
03The AI's high-attention cells coincided with those identified as diagnostically relevant by human experts.
04The algorithm could highlight subtype-specific cells and deconvolve blood smear composition without requiring single-cell annotation during training.

Application

Design takeaway

How to apply

In developing automated diagnostic tools, implement methods that not only provide accurate predictions but also clearly indicate the features or data points that led to those predictions, allowing for expert review and validation.

Project actions

01Consider using AI models that offer interpretability features.
02When evaluating AI performance, compare its decision-making process to that of human experts in the field.
03Think about how to present AI findings in a way that builds confidence with the end-user.

Method & Evidence

AimCan an explainable AI model accurately classify genetic subtypes of acute myeloid leukemia (AML) from blood smear images by identifying diagnostically relevant cells, and does its focus align with human expert identification?

MethodMachine Learning (Explainable AI, Multiple Instance Learning)

ProcedureA single-cell based explainable multiple instance learning algorithm (SCEMILA) was trained on over 80,000 single white blood cell images from digitized blood smears of 129 AML patients (four genetic subtypes) and 60 healthy controls. The model's attention mechanisms were analyzed to compare its focus on diagnostically relevant cells with that of human experts. Performance was evaluated using F1 scores and discrimination accuracy.

Sample189 patients (129 AML, 60 controls)

ContextMedical diagnostics, Hematology, Digital Pathology

Variables

IVBlood smear images (features of white blood cells)

DVClassification of AML subtypes (including healthy controls)

CVImage quality, patient demographics (implicitly), WHO-defined AML subtypes, diagnostic criteria used by human experts.

Strengths & Limitations

Strengths

+Use of a large, real-world dataset.
+Demonstration of explainability aligning with expert knowledge.
+Development of a publicly available dataset and tool.

Limitations

The AI's performance might vary with different image qualities or variations in sample preparation. The dataset, while large, represents a specific patient population, and real-world deployment might encounter more diverse cases.

Reliability & validity

Reliability is supported by the use of cross-validation and reporting of mean ± s.d. Validity is addressed by the concordance between AI's attention and expert identification, and the high classification accuracy for AML subtypes.

Think critically

How can the 'explainability' of an AI be objectively measured and validated to ensure it truly reflects expert reasoning, rather than just highlighting prominent features?

Design Principles

"Diagnostic AI systems should be designed with inherent explainability to validate their decision-making processes against established domain expertise."

This research demonstrates the potential for AI to augment diagnostic processes in healthcare, offering consistent and potentially faster analysis. The explainability of the AI builds trust and allows for validation against established expert knowledge, crucial for adoption in clinical settings.

What This Means for Your Design

This study shows that a smart computer program can look at blood cell pictures and tell if someone has a certain type of blood cancer, just like a doctor would. It's good because the program can explain *why* it made its decision by pointing out the specific cells it looked at, which makes doctors more likely to trust it.

How to use in your project

1.Reference this study when discussing the importance of explainable AI in your design project, especially if your project involves complex data analysis or decision support.
2.Use the findings to justify the need for transparency in your AI-driven design solutions.

Add to My Project

Quick Cite

Paragraph starter

The development of explainable AI, as demonstrated by SCEMILA in classifying acute myeloid leukemia subtypes (Hehr et al., 2023), highlights the critical need for transparency in automated decision-support systems. By aligning AI's focus with expert human identification, such models can foster trust and facilitate adoption in complex diagnostic workflows, suggesting that design projects involving AI should prioritize interpretability alongside accuracy.

Source

PLOS Digital Health

Explainable AI identifies diagnostic cells of genetic AML subtypes

journal · 2023

View source

Questions About This Research

What does the research say about ai-driven cell analysis achieves 94% accuracy in aml subtype diagnosis, matching expert pathologist focus?: When developing AI for diagnostic applications, prioritize explainability to ensure transparency and alignment with expert knowledge, thereby fostering user trust and facilitating clinical integration. Evidence: PLOS Digital Health (2023).
Why does "AI-driven cell analysis achieves 94% accuracy in AML subtype diagnosis, matching expert pathologist focus" matter for design?: This research demonstrates the potential for AI to augment diagnostic processes in healthcare, offering consistent and potentially faster analysis. The explainability of the AI builds trust and allows for validation against established expert knowledge, crucial for adoption in clinical settings.
How can designers apply this research?: When developing AI for diagnostic applications, prioritize explainability to ensure transparency and alignment with expert knowledge, thereby fostering user trust and facilitating clinical integration.
What were the main findings?: SCEMILA perfectly discriminated between AML patients and healthy controls.. The model detected the APL subtype with an F1 score of 0.86 ± 0.05.. The AI's high-attention cells coincided with those identified as diagnostically relevant by human experts.. The algorithm could highlight subtype-specific cells and deconvolve blood smear composition without requiring single-cell annotation during training.
What research method was used?: Machine Learning (Explainable AI, Multiple Instance Learning) with 189 patients (129 AML, 60 controls).
How strong is the evidence?: Evidence strength is rated Strong effect, based on a 2023 journal from PLOS Digital Health.
What should I do differently in my next project?: In developing automated diagnostic tools, implement methods that not only provide accurate predictions but also clearly indicate the features or data points that led to those predictions, allowing for expert review and validation.
What are the limitations?: The study focused on specific AML subtypes and blood smear analysis; generalizability to other diseases or sample types would require further validation. The performance metrics are based on a specific dataset and cross-validation strategy.