Profile Generators: A Link between the Narrative and the Binary Matrix Representation

Raoul H. Kutil, Georg Zimmermann, Barbara Strasser-Kirchweger, Christian BorgeltarXiv 2025 · DOI · arXiv

Introduction

In the field of computational psychiatry, researchers face a "combinatorial nightmare": mathematically comparing mental health disorders when a single diagnosis can manifest in billions of different ways. A recent study by researchers at the University of Salzburg and Paracelsus Medical University introduces Profile Generators, a framework designed to translate narrative diagnostic criteria into machine-readable logic, drastically reducing the computational complexity of disorder comparison.

The Challenge of Narrative Criteria

Diagnostic manuals like the DSM-5 describe mental disorders using narrative rules, such as "at least five of the following nine symptoms" or "at least one of symptoms A, B, or C must be present." While these rules are intuitive for clinicians, they are difficult for computers to process directly.

Annotated DSM-5 criteria for Schizophrenia, highlighting the components of diagnostic logic.

To analyze similarity, researchers often use a "Binary Matrix Representation" (BM), where each row represents a valid symptom combination (a profile). However, for complex disorders, the number of valid profiles explodes. Major Depressive Disorder (MDD) has over 1.3 billion valid combinations, while Panic Disorder has over 3 billion. A brute-force pairwise comparison between these two would require roughly 4 quintillion calculations ($4 \times 10^{18}$), making it computationally infeasible even for modern supercomputers.

Formalizing Diagnostic Logic

The authors propose bridging this gap by formalizing the narrative text into strict set-theory operations. They define symptoms, profiles, and criteria mathematically to allow for the creation of "generators"—predefined functions that encode diagnostic logic.

Mathematical definitions bridging narrative symptoms to formal set-theory representations.

The framework utilizes five types of generators (G0–G4) to handle varying levels of complexity, from simple identity functions to "Union Product" functions that manage nested "and/or" logic. This allows the DSM-5 to be stored as a compact list of rules rather than an unmanageable matrix.

Optimization through Conditional Generators

The core innovation of the paper is the Conditional Generator. Instead of generating every possible symptom profile and then searching for the most similar pair, the authors use set theory to identify the specific profiles that maximize similarity. By segmenting symptoms into those shared by both disorders and those unique to one, the generator is "forced" to produce only the most overlapping profiles. This approach treats the diagnostic criteria as a search space that can be pruned before a single profile is even created.

Results and Clinical Validation

The efficiency gains from this method are transformative. For the comparison of MDD versus Panic Disorder, the number of required cosine similarity calculations was reduced from 4.29 quintillion to exactly one. Calculations that would have taken years were completed in seconds on standard hardware.

Furthermore, the resulting Maximum Pairwise Cosine Similarity (MPCS) scores aligned with clinical expectations. For instance, MDD and Persistent Depressive Disorder (PDD) showed high similarity (0.763), while MDD and Panic Disorder showed much lower overlap (0.139).

Conclusion and Future Work

While the framework successfully optimizes for maximum similarity, calculating average similarity still requires accounting for all possible profile pairs. Future work aims to automate the translation of narrative criteria using Natural Language Processing (NLP) and expand the application of Profile Generators to other medical fields involving complex, rule-based diagnostic pathways.

Summary updated

BibTeX
@misc{kutil2025profile,
  author = {Raoul H. Kutil and Georg Zimmermann and Barbara Strasser-Kirchweger and Christian Borgelt},
  title = {{Profile Generators: A Link between the Narrative and the Binary Matrix Representation}},
  year = {2025},
  eprint = {2511.19506},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.19506},
}
Figure