HeyDonto is an applied-research company. Every product we ship — Axiomera, Quantara Health AI, and Conduit — is built on mathematics that has cleared peer review. This page catalogs the 14 publications and 11+ patents pending that define our scientific foundation: quantum-inspired semantic intelligence, federated clinical learning, metabolic oncology, and FHIR-native harmonization at population scale.
The overwhelming majority of healthcare AI today is trained on data that no one has taken the time to make mean the same thing. CDT code D2750 arrives as CRN-PFM, crown_pfm, or a free-text string depending on which practice management system sent it. Patient identity, units of measure, laterality, and temporal ordering are rarely consistent even inside a single health system.
Our research begins where most vendor roadmaps end. Rather than fine-tuning another downstream model, our team has spent the last two years publishing the formal machinery for a semantic layer that sits before analytics, reasoning, and AI: a Hilbert-space representation of clinical concepts; evolutionary optimization of the networks that classify them; federated learning protocols that preserve privacy across 47 institutions; and FHIR-native overlays that make two conformant endpoints actually interoperable.
We don't build applications on top of unreliable data. We build the math that makes the data reliable in the first place.
This matters because healthcare is a regulated industry. Determinism, auditability, and fairness are not features — they are licensure requirements. The papers catalogued below establish the information-theoretic bounds on harmonization fidelity, the fairness-preserving constraints that protect demographic subgroups, and the statistical validation protocols that turn AI claims into clinical evidence.
For partners and investors: this is what we mean when we say "infrastructure-tier." These publications are the artifacts that make Axiomera, Quantara, and Conduit defensible against competitors who marketed first and will justify later.
Foundational AI architecture and semantic infrastructure: evolutionary neural networks, quantum-inspired self-healing frameworks, multi-scale neuromorphic integration, federated readmission prediction, graph-based concept extraction, quantum-superposition harmonization, and FHIR-native profile governance.
Translational cancer research with therapeutic and diagnostic implications: lncRNA-driven metabolic reprogramming in hepatocellular carcinoma, epigenetic drug-resistance reversal in advanced non-small cell lung cancer, and deep-learning radiomics for preoperative glioma grading across five tertiary neurosurgical centers.
The bridge layer: harmonization-aware forecasting of pharmaceutical need and utilization across multi-site federated data, and the unified mathematical framework for semantic classification that underlies both products. These papers connect our harmonization machinery to concrete decision objects — a drug, a therapy class, a policy change.
Nine publications establish the mathematical and architectural foundation for Axiomera's Semantic Classification Engine (SCE), Data Mapper Model (DMM), and Data Harmonization Model (DHM). Together they span the full stack — from the evolutionary neural networks that optimize classification, to the quantum-inspired frameworks that keep candidate interpretations in superposition, to the FHIR-native governance lifecycle that makes harmonization artifacts auditable.
Introduces the foundational technique now embedded in Axiomera's classification layer: evolutionary operators — natural selection, model combination, and random weight mutations — are used to jointly optimize CNN and RNN architectures for large-scale digital data. The paper establishes measurable improvements in accuracy, convergence speed, and generalization versus gradient-based baselines, and provides the formalism behind our Evolutionary Neural Network patent applications (63/675,138 and 19/181,522).
Proposes a self-healing framework combining quantum-inspired optimization (candidate solutions held in superposition), biomimetic error detection modeled on biological immune response, and fractal cross-architectural propagation. Validated across 15,000 software engineering tasks in five domains. The principles established here — superposition of candidate interpretations, automated error correction, fractal scalability — are directly inherited by Axiomera's Active Control layer.
A hierarchical five-layer architecture that unifies neuromorphic spiking substrates, quantum-inspired parallelism, and fractal data processing. The framework demonstrates genuine emergent properties — self-organization, adaptive specialization, distributed error correction — without explicit programming. Forms the theoretical backbone for Axiomera's ability to scale semantic reasoning across six orders of magnitude of data volume.
Introduces dynamic cognitive substrates that eliminate the traditional distinction between training and inference — the system continuously absorbs real-time streams, autonomously restructures itself, and prioritizes valuable information through biologically-inspired memory consolidation. This is the scientific basis for Axiomera's ability to improve without retraining cycles, and for our claim that the platform learns from what it's already seen.
The largest prospective federated-learning study in our catalog. Histogram-based XGBoost deployed across 47 U.S. healthcare institutions in a real-world quality improvement initiative, with MIMIC-IV external benchmarking. Establishes that federated learning is practical — not theoretical — at multi-institutional scale, with amplified differential privacy (ε = 0.68, δ = 10⁻⁵) and 99.1% operational uptime.
Unified framework integrating hybrid-NLP graph concept extraction, statistically validated relationship discovery (with multiple-testing correction), and temporal pattern analysis via bidirectional LSTMs with clinical attention. Evaluated across three healthcare systems comprising 492,542 encounters from 127,834 patients. This is the published evidence that the Axiomera Semantic Evidence Graph (SEG) materially improves diagnostic coding, comorbidity detection, and decision-support relevance.
Treats multi-institutional EHR data as a quantum-inspired superposition of clinical realities. Drawing from the Many-Worlds interpretation, fragmented patient data is modeled as branches of a global clinical wavefunction evolving within a Hilbert space of semantic states. Harmonization is formulated as a controlled transition from high-entropy superposition to approximately classical, clinically consistent branches via decoherence-like mechanisms constrained by medical knowledge, fairness rules, and governance. Provides the mathematically rigorous language for weighting and harmonizing incompatible clinical data at population scale.
The most direct theoretical statement of the Semantic Classification Engine (SCE). Clinical concepts are vectors in a Hilbert space H with basis states corresponding to codes from ICD-10-CM, SNOMED CT, LOINC, and CDT. Classification is formalized as a quantum-inspired measurement operator yielding calibrated probability distributions over semantic types. The paper proves convergence of the classification algorithm and establishes ε-consistency conditions — the formal guarantees that Axiomera offers to regulated partners.
HL7 FHIR solves syntactic exchange but not semantic equivalence — two endpoints can be FHIR-compliant while encoding the same concept using different profiles, coding systems, and extensions. This paper proposes a vendor-neutral overlay with three services: a federated identity provider, an observed-profile analyzer, and a governed harmonization layer. The key primitive is the Data Profile Identifier (DPID) — a content-addressed fingerprint enabling decentralized profile comparison, clustering, and drift detection. This is the paper that makes Conduit's four exchange patterns mathematically rigorous.
Three translational oncology papers establish Quantara Health AI's scientific credibility as a cancer-intelligence subsidiary. Each describes a distinct therapeutic or diagnostic mechanism — a novel lncRNA target in hepatocellular carcinoma, a dual epigenetic compound reversing chemoresistance in non-small cell lung cancer, and a habitat-enhanced deep-learning framework for preoperative glioma grading — with direct clinical translation paths.
Identifies HELC (Hepatic Enhancer of Lipid Carcinogenesis), a previously uncharacterized long non-coding RNA upregulated in HCC tissues across multiple independent cohorts (n=312, fold-change 7.23). HELC scaffolds the AP1 transcription factor complex at the FASN promoter, orchestrating de novo lipogenesis. The paper also introduces HELCi-7, a first-in-class small molecule inhibitor disrupting the HELC-AP1 interface — the therapeutic candidate that anchors Quantara's hepatocellular-carcinoma franchise.
Integrated multi-omics (scATAC-seq, scRNA-seq, proteomics) across treatment-naive and drug-resistant NSCLC models identifies 847 differentially accessible regulatory regions organized into 23 coordinated enhancer clusters. Introduces DT-847, a rationally-designed compound with nanomolar binding affinity for both DNMT1/3A and HDAC1/2 — dual targeting produces metabolic collapse and restores chemosensitivity. The accompanying Epigenetic Resistance Index (ERI) is the diagnostic companion now being scoped into Quantara's precision-oncology platform.
Multi-center diagnostic study across five tertiary neurosurgical centers (n=1,060) combining radiomics feature extraction, deep-learning representation learning, and a novel tumor habitat analysis capturing spatial heterogeneity patterns that correlate with underlying biology. Habitat-enhanced fusion achieves AUC 0.946 for high-grade vs low-grade discrimination and exceeds AUC 0.86 for all evaluated molecular markers — meaningful lift over radiomics or DL alone. Prospective evaluation shows the model shifted management decisions in 42% of cases.
The bridge layer connects Axiomera's harmonization machinery to concrete pharmaceutical and policy decisions. This is where our semantic infrastructure earns its keep — by producing forecasts that explicitly separate clinical need from realized utilization, preserving uncertainty end-to-end, and enabling federated training across sites that cannot share raw data.
Introduces the first framework that leverages uncertainty-preserving clinical harmonization to forecast the intersection between a harmonized population state and an external decision object — a drug, a therapy class, a policy change. Critically, the framework explicitly separates forecasts of clinical need from realized utilization — a distinction that matters when access barriers, supply chain issues, or substitution patterns are decoupled from underlying demand. Builds on Papers 5, 6, and 7. Maintains calibration and fairness across demographic subgroups under federated settings.
Our research is not bench-only. Every framework above has been validated against real-world deployments, external benchmarks, or independent cohorts — then documented in a third-party External Validation Report (February 2026) confirming the semantic-first architecture across healthcare, enterprise data, and federated-learning scenarios.
Our patent strategy mirrors our research strategy: publish first, then protect the engineered applications. Current filings span US, European, and PCT jurisdictions, with non-provisional conversions tracked to the November 2026 deadline. Patent counsel is Jeffrey Sheriff of Pillsbury Winthrop Shaw Pittman LLP. These patents do not attempt to claim mathematical laws of nature — they claim the specific engineered systems, architectures, and methods that make the underlying science deployable in regulated industries.
Quantum-inspired measurement operators over Hilbert-space representations of clinical concepts bound to ICD-10-CM, SNOMED CT, LOINC, and CDT basis states, with calibrated confidence estimates and Semantic Evidence Graph output.
Evidence-based mapping of heterogeneous source fields to canonical target schemas, with uncertainty preservation and governed versioning across source systems including Eaglesoft, Open Dental, Dentrix, Epic Clarity, and Cerner.
Methods for transforming multi-source clinical records into harmonized bundles with per-field provenance, fairness-preserving constraints, and auditable rollout of normalization artifacts.
CNN / LSTM architectures optimized via evolutionary operators — natural selection, model combination, random weight mutations — for large-scale digital data processing. The foundational technique underlying Axiomera's classification layer.
Nine claims covering topological soliton detection, renormalization-group scene parsing, and manifold optimization — engineered applications of discovered DFT mathematics for aerospace, defense, and robotics verticals.
Ten claims across four anomaly-detection channels, applicable to aviation inspection (TEC-USA alignment), industrial monitoring, and financial fraud. Extends the DFT patent family into adjacent safety-critical verticals.
Many-Worlds harmonization methods: modeling fragmented patient data as branches of a global clinical wavefunction and performing controlled decoherence-like transitions constrained by medical knowledge and governance rules.
Content-addressed fingerprints of stable FHIR profile signatures enabling decentralized profile comparison, clustering, and drift detection across organizations without prior coordination. Powers Conduit's federated identity services.
Histogram-based XGBoost federated training protocol with amplified differential privacy guarantees, site-compute optimization, and fairness monitoring — validated across 47 U.S. healthcare institutions.
Multi-omic biomarker signature combining scATAC-seq, scRNA-seq, and proteomic features to predict chemotherapy resistance in advanced NSCLC — 94.7% sensitivity, 91.2% specificity, AUC 0.927. Quantara translational IP.
Small-molecule inhibitor disrupting the HELC–AP1 transcription-factor interface, demonstrating 72% tumor growth inhibition in patient-derived HCC xenograft models without observable toxicity.
Spatial heterogeneity decomposition methods integrated with deep-learning fusion architectures for preoperative glioma grading and molecular-marker prediction across multi-center MRI data.
Partners, investors, and regulatory counsel can request the full library of peer-reviewed publications, patent summaries, and the February 2026 External Validation Report. Technical deep-dives are available on request.