Herbotuple

diagnostics de la maladie => AI => Recettes à base de plantes

Given the disease diagnosis, Herbotuple outputs herbal recipe for the disease.

Disease diagnosis is in ICD-9 code, and every input can have up to three ICD-9 codes. In response to the input disease and comorbidity, Herbotuple outputs an herbal recipe, including the doses in grams of the herbs and formulas in the recipe. Medication schedule of the recipe is also included in the output. The herbs and formulas in the recipe have been widely used by Chinese people and are in the form of concentrated extract granules, which are manufactured by GMP-certified pharmaceutical companies and are available in most regions of the world.

The algorithm behind Herbotuple is artificial intelligence programs developed by Google Brain. The data for the AI deep learning consists of 340 million de-identified records of disease diagnoses and herbal prescriptions from the health insurance reimbursement claims, made in the period between 2004 and 2013, in the National Health Insurance Research Database, Ministry of Health and Welfare, Taiwan. The ICD-9 codes together with patient's age and sex were considered as a sentence in one language; and the prescription composed of herbs and formulas as a sentence in the other language. We then applied Google's latest neural translation algorithm to the 340 million sentence pairs. Prior to deep learning, an important sampling method using medoids was developed to deal with the issue of imbalanced disease-age-sex classes in the raw data. Another tweak involved modeling the weight distribution of the herbs and formulas making up the recipe by Zipf's law. A Zipf's exponent was extracted to represent a sentence's 'tone'. Training of the recurrent neural network proceeded until the translation achieved a word perplexity of 7. Case studies of input and output showed that Herbotuple performed excellent translation.