Machine learning in homeopathy

Machine learning is a class of statistical pattern-matching tools that learn associations from labelled or co-occurring data. In homeopathy software it occupies a narrow, useful place: the parts of repertorisation that are search-and-rank problems rather than clinical-judgement problems. A semantic-search ranker that returns the rubric Mind — Fear, dark, of when a practitioner types "scared of being alone at night" is doing string-to-rubric retrieval. It is not adjudicating the choice of Stramonium for the patient. That distinction is load-bearing throughout what follows.

What machine learning actually does

Modern machine learning in clinical-software contexts is dominated by two families: supervised classifiers trained on labelled examples, and unsupervised embedding models that project text, images or audio into a vector space where semantic neighbours cluster together. Both are pattern-finders, not reasoners. A model trained on millions of (sentence, rubric) pairs from the Complete Repertory can learn that "headache better lying down" sits near Head — Pain, lying, amel, but it has no opinion on whether the matched rubric is the correct repertorial entry for the case in front of the practitioner.

Word and sentence embeddings — the architecture under most semantic search today — date to the word2vec family of models and were extended to context-sensitive embeddings with the BERT line of work. Repertory-side application is straightforward: every rubric is encoded once into a vector, every practitioner query is encoded at search time, and the top-k cosine-nearest rubrics are returned. The interpretability question — why the model ranked rubric X above rubric Y — is open, and a retrieval suggestion never absolves the practitioner from cross-checking against the materia medica.

Calibration of a rubric-retrieval model is constrained by the training corpus. A model trained mostly on modern English consultation notes will under-represent the nineteenth-century language of Kent and Hering, which means a paragraph-scale historical case description can land far from its closest classical rubric in vector space. Any ML-assisted repertorisation tool used clinically should disclose its training-data scope and its model version, so a retrieval result can be read with the appropriate provenance.

Classical algorithms versus machine-learning approaches

The classical repertorisation algorithm — descended from Boenninghausen's Therapeutic Pocket Book and formalised by Kent — is a counting procedure over a practitioner-curated rubric list. Machine learning does not replace that procedure. It changes the front end: how rubrics are found, how case notes are converted into rubric candidates, and how audio or photographic data is translated into searchable text. The analysis table that displays remedy scores after rubrics are added remains a Boolean-style aggregation, not a model prediction.

Where machine learning genuinely helps

Three workflow gates dominate practitioner time in contemporary case-taking software, and each is amenable to a different machine-learning component.

• Semantic search of the repertory. Embedding-based search lets you type symptoms in patient language ("fear of the dark", "headache better lying down") and surface relevant rubrics without remembering exact repertory wording. The cleanest interfaces toggle Semantic Search alongside Keyword Search per query.
• Notes-to-rubrics extraction. A large language model reads consultation notes and proposes rubrics, designed to catch Strange, Rare and Peculiar symptoms that scroll past the eye on a long case. The proposed rubrics are suggestions to accept or reject, never auto-inserted into a final analysis.
• Live audio transcription with rubric suggestion. Speech-to-text plus an SRP extractor produces a running transcript and a candidate rubric list during the consultation itself. A well-built implementation discards the raw audio after the session and retains only the transcript and a structured summary.

Each of these is a retrieval and translation task. None of them adjudicates whether the remedy that finishes top of the analysis table will help the patient. The practitioner remains the clinical agent; the model is a faster index.

Where machine learning stops and clinical reasoning begins

A model trained on rubric–remedy co-occurrence answers a retrieval question: which remedy does the canon associate with this symptom set. That is not the clinical question you bring to the case, which is whether this remedy, for this patient, at this moment, fits the totality as the tradition teaches. No ranking algorithm replaces that judgement, because remedy selection is relational and individual in ways pattern-matching over aggregated data does not capture.

A model result is a retrieval signal, not a verdict on the remedy. The work is to interrogate it against the materia medica and the case.

In clinical use, record per case which rubrics were model-suggested versus practitioner-entered, and the model version that produced the suggestion. That audit trail is what lets a later reviewer separate the practitioner's reasoning from the model's retrieval, and it is the discipline responsible use of the technology requires.

Open questions worth testing

An embedding model trained jointly on the Complete Repertory and the chronologically-ordered canon — Hahnemann's Organon and Chronic Diseases, Kent's Lectures and Repertory, Hering's Guiding Symptoms, Boger's Synoptic Key, Vithoulkas's Science of Homeopathy, Sankaran's sensation-method volumes — may surface remedies that modern Boolean repertorisation underweights, because their characterising symptoms are paragraph-scale rather than rubric-scale. The question is testable: hold out a set of paragraph-scale historical case notes, encode them, retrieve top-k remedies, and ask a panel of practitioners to grade fidelity.

A second open question concerns variance: whether transcription-plus-extraction narrows or widens rubric selection across practitioners working from the same audio. Model suggestions could pull novice and expert toward a shared median, or anchor both on a model artefact and suppress the diversity of reasoning. Both deserve real studies before strong claims are made in either direction.

Practitioners who want to audit how a contemporary stack handles these workflows can run a real case through semantic search, notes-to-rubrics and live audio mode in a working implementation and judge each output alongside the canon.

References

LeCun, Y., Bengio, Y. and Hinton, G. (2015) "Deep learning," Nature 521(7553):436–444, https://www.nature.com/articles/nature14539.

Mikolov, T., Chen, K., Corrado, G. and Dean, J. (2013) "Efficient estimation of word representations in vector space," arXiv:1301.3781, https://arxiv.org/abs/1301.3781.

Devlin, J., Chang, M.-W., Lee, K. and Toutanova, K. (2019) "BERT: Pre-training of deep bidirectional transformers for language understanding," Proceedings of NAACL-HLT, https://aclanthology.org/N19-1423/.

Hahnemann, S., Organon of the Medical Art and The Chronic Diseases; Kent, J. T., Lectures on Homoeopathic Materia Medica and Repertory of the Homoeopathic Materia Medica; Hering, C., The Guiding Symptoms of our Materia Medica; Boenninghausen, C. von, Therapeutic Pocket Book; Boger, C. M., A Synoptic Key of the Materia Medica; Vithoulkas, G., The Science of Homeopathy; Sankaran, R., The Substance of Homoeopathy and The Sensation in Homoeopathy.