Science & validation

Biology-guided, explainable AI

Cancer biology cannot be explained by a single biomarker. Harae Dx’s AI learns the relationships between five complementary biological pathways to generate an explainable cancer risk prediction.

The Harae Dx AI Platform

From biological data to clinically actionable insights

A multi-layered, explainable AI architecture that decodes protein relationships to deliver accurate, early cancer risk prediction.

1

Clinical Data Layer

Integrated clinical context — history, demographics, imaging, and patient risk factors — enhances biological interpretation.

2

Protein Biomarker Layer

Quantitative measurement of five complementary protein biomarkers from a simple blood sample, spanning cell signaling, matrix remodeling, tumor-associated, inflammatory, and angiogenic pathways.

3

Biology Knowledge Layer

A biology-guided knowledge base captures the pathways and relationships that drive cancer development and progression.

4

AI Intelligence Layer

A proprietary AI engine learns complex biomarker relationships and patterns using advanced machine learning and explainable AI.

5

Clinical Decision Engine

AI insights are translated into risk stratification with calibrated probabilities and recommendations.

6

Output & Actionable Insights

Clinicians receive clear, actionable results in under 30 minutes to guide timely decisions.

The AI flywheel

Every patient makes the platform smarter

A continuously learning system that improves with every patient, in every setting, across every geography.

1

More Patients Tested

Across hospitals, clinics, and countries.

2

Proprietary Clinical Knowledge

Biomarker data linked with clinical outcomes, demographics, imaging, and follow-up.

3

Biological Pattern Learning

The AI learns biomarker interactions and disease signatures, and improves explainability.

4

Better Clinical Intelligence

Higher sensitivity, higher specificity, earlier detection, and explainable recommendations.

5

Greater Clinical Adoption

More hospitals, more clinicians, more trust, more usage.

6

Expanding Biological Intelligence

New populations, geographies, cancer types, and biomarkers.

Model development & validation

Rigorously trained, then independently validated

Before any human sample was tested, the underlying AI risk model was built and stress-tested computationally, then confirmed against real biological samples in independent retrospective studies.

In silico model training & testing

Reflects the current five-biomarker AI model.

5,000+

In silico patient profiles used to train the model

1,000+

In silico patient profiles used to test the model

95%+

Sensitivity and specificity achieved in silico

The model was then checked against 350+ actual patient samples obtained from research repositories in the US and Europe, ahead of the independent retrospective studies below.

Note on biomarker panel: the retrospective studies below (Study 1, Study 2, and the seven additional lab studies) were conducted using Harae Dx’s original four-biomarker panel. A fifth, angiogenic biomarker was subsequently added to build the current five-biomarker AI model reflected in the in silico results above; retrospective clinical validation of the five-biomarker panel is ongoing.

Independent retrospective studies

Conducted using the original four-biomarker panel.

StudyPlasma samplesSensitivitySpecificity
Independent research institution (U.S.)120 (100 with cancer, 20 controls)91.67%100%
External clinical validation partner92 (68 with cancer, 24 controls)97.06%100%

Broader lab validation: 7 additional retrospective studies

Beyond the two studies above, the same four-biomarker panel was further validated across 7 independent lab studies using retrospective plasma samples — ranging from 56 to 308 samples each, with performance consistent across sample sets and independent of patient ethnicity.

StudySample sizeSensitivitySpecificityAccuracy
Study 375 (44 with cancer, 31 controls)92.11%85.71%89.39%
Study 492 (54 with cancer, 38 controls)82.50%80.00%81.43%
Study 5120 (97 with cancer, 23 controls)92.50%85.71%91.95%
Study 656 (34 with cancer, 22 controls)85.19%81.25%83.72%
Study 782 (41 with cancer, 41 controls)94.74%100.00%97.40%
Study 8110 (52 with cancer, 58 controls)100.00%100.00%100.00%
Study 9308 (140 with cancer, 168 controls)89.92%94.41%92.37%

Across these 7 studies: average sensitivity and specificity of approximately 90%, average positive predictive value of 93.50%, average negative predictive value of 82.82%, and average accuracy of 90.89%.

Performance figures reflect retrospective, preliminary validation studies conducted with external research and clinical partners. Sensitivity in these studies was independent of cancer stage. MammoAlert is not yet FDA cleared and results reported here are for informational purposes; the product is for investigational use as clinical validation continues.
Regulatory progress

Building toward a global regulatory pathway

FDA — Breakthrough Device Designation

Harae Dx has received FDA Breakthrough Device Designation, supporting an expedited path toward regulatory review.

India — fast-track clinical pathway

Fast-track clinical trial pathway discussions underway in India.

“Most current cancer diagnostics are designed for patients who have already entered the healthcare system. Harae Dx is designed to expand access to the much larger population that is not routinely screened.”