Medical diagnosis generation
Leveraging generative artificial intelligence to provide diagnoses for patients. In some examples, elimination rules based on patient intake data are applied to a very large corpus of candidate diagnoses, such as all known human medical diagnoses, to narrow down the very large corpus of candidates to a meaningful differential diagnosis or final diagnosis. In some examples, targeted information is elicited to determine and apply additional elimination rules to further narrow down the candidate diagnoses. In some examples, patient intake data is classified into risk data and symptom data and processed by data type-specific machine learning models. In some examples, patient intake data is reconciled and correlated between risk and symptom data by specific machine-learning models to improve diagnostic clustering and accuracy.
1 . A computer-implemented method for generating a diagnosis of a human disease of a patient, comprising:
receiving unstructured data, including medical journals;
cleaning the unstructured data to generate cleaned data;
converting the cleaned data into a standardized format, the standardized format including a table having rows and columns;
training at least one machine learning model using the table to generate at least one trained machine learning model;
receiving, with the at least one trained machine learning model and via an input device, patient intake data related to the patient;
accessing, with the at least one trained machine learning model and from at least one database, a collection of at least 7,000 candidate human medical diagnoses;
determining, by the at least one trained machine learning model, a plurality of elimination rules for a candidate diagnosis elimination process, wherein the plurality of elimination rules are based on the patient intake data, wherein the at least one trained machine learning model includes a Process Reward Model (PRM) combined with a Tree of Thoughts (TOT) method of reasoning including a tree, the PRM being configured to assess and score each step of the candidate diagnosis elimination process, and the TOT method being configured to navigate through a plurality of reasoning steps by adjusting a route through branches of the tree based on both prior conclusions and future assertions to identify a possible path to the diagnosis, the possible path including requesting and receiving additional patient intake data to rule out at least one candidate diagnosis;
applying, by the at least one trained machine learning model, the plurality of elimination rules to the collection to generate a subset of candidate human medical diagnoses from among the collection;
generating the diagnosis based on the subset, the diagnosis including one of the medical diagnoses from among the subset; and
outputting, using the at least one trained machine learning model, to a user interface of an output device:
the diagnosis; and
an explanation for the diagnosis, the explanation including both symptom data and non- symptom risk data obtained from the patient intake data.
2 . The computer-implemented method of claim 1 , wherein the collection includes at least 28,000 candidate human medical diagnoses.
3 . The computer-implemented method of claim 1 , wherein the plurality of elimination rules are determined based on factors of the patient derived from the patient intake data, the factors including one or more of: age, gender, race, a blood test result, a urine test result, a genetic test result, an absent radiographic finding, a present radiographic finding, an ultrasound result, a biopsy result, an absent physical sign finding, a present physical sign finding, an absent symptom, a present symptom, an absent personal medical history, a present personal medical history, an absent history of exposures, a present history of exposures, an absent history of traumas, a present history of traumas, a present history of pregnancy, a family history, an electrocardiogram result, an electroencephalogram result, a negative peripheral smear, positive peripheral smear, a negative culture, a positive culture, a drug history, a drug allergy, a drug side effect, a drug contraindication, an interaction between drugs, an interaction between a drug and food, an absent patient risk factor, a present patient risk factor, socio-ethnic- economic classification, and geographic location.
4 . The computer-implemented method of claim 1 , further comprising:
determining, by the at least one trained machine learning model, at least one inclusion rule based on the patient intake data; and
excluding, by the at least trained one machine learning model, one of the plurality of elimination rules based on the at least one inclusion rule such that the one of the plurality of elimination rules is not applied by the at least one trained machine learning model to the collection.
5 . The computer-implemented method of claim 1 , further comprising:
determining, by the at least one trained machine learning model, an insufficiency in the plurality of elimination rules;
generating, with the at least one trained machine learning model and based on the patient intake data, at least one question to pose to the patient or at least one test to administer to the patient;
receiving, with the at least one trained machine learning model, the additional patient intake data based on at least one answer to the at least one question or at least one result of the at least one test; and
determining, by the at least one trained machine learning model, at least one additional elimination rule based on the additional patient intake data, wherein the applying includes applying, by the at least one trained machine learning model, the at least one additional elimination rule to the collection to generate the subset of candidate human medical diagnoses from among the collection.
6 . The computer-implemented method of claim 1 , wherein the patient intake data is elicited from the patient with a virtual assistant that converses with the patient and verbally asks the patient questions relating to one or more physical symptoms and to a family medical history of the patient.
7 . The computer-implemented method of claim 1 , wherein the patient intake data is obtained by one or more of:
a microphone that captures sounds generated by the patient;
a visual camera that captures images of the patient;
a thermometer that measures a body temperature of the patient;
a blood pressure meter that measures a blood pressure of the patient;
an oximeter that measures an oxygen concentration in blood of the patient; and
a pulse meter that measures a heart rate of the patient.
8 . The computer-implemented method of claim 1 , further comprising:
outputting, by at least one trained machine learning model and based on the patient intake data, a recommendation to deprescribe a medication.
9 . A computer-implemented method for generating a diagnosis of a human disease of a patient, comprising:
receiving unstructured data, including medical journals;
cleaning the unstructured data to generate cleaned data;
converting the cleaned data into a standardized format, the standardized format including a table having rows and columns;
training a plurality of machine learning models using the table to generate a plurality of trained machine learning models;
receiving, via an input device, patient intake data;
determining that a first subset of the patient intake data relates to one or more physical symptoms of the patient;
determining that a second subset of the patient intake data relates to a predisposition of the patient to each of a plurality of diseases independently of the one or more physical symptoms;
providing the first subset of the patient intake data and the second subset of the patient intake data as inputs to the plurality of trained machine learning models, including:
a risk model that receives, as risk input, the second subset of the patient intake data and generates, as risk output, the predisposition of the patient to the one or more diseases; and
a diagnosis model that receives, as diagnosis input, the first subset of data and the risk output, wherein the diagnosis model includes a Process Reward Model (PRM) combined with a Tree of Thoughts (TOT) method of reasoning including a tree, the PRM being configured to assess and score each step of a candidate diagnosis elimination process, and the TOT method being configured to navigate through a plurality of reasoning steps by adjusting a route through branches of the tree based on both prior conclusions and future assertions to identify a possible path to the diagnosis, the possible path including requesting and receiving additional patient intake data to rule out at least one candidate diagnosis;
generating, with the plurality of trained machine learning models, the diagnosis for the patient based on the first subset of the patient intake data and the second subset of the patient intake data, the generating including determining, by the diagnosis model, a plurality of elimination rules for the candidate diagnosis elimination process, wherein the plurality of elimination rules are based on the patient intake data; and
outputting, by the diagnosis model, to a user interface of an output device:
the diagnosis; and
an explanation for the diagnosis, the explanation including both symptom data and non-symptom risk data obtained from the patient intake data.
10 . The computer-implemented method of claim 9 , further comprising:
outputting, by at least one of the plurality of trained machine learning models and based on the first subset of the patient intake data and the second subset of the patient intake data, a recommendation for one or more tests to administer to the patient to confirm or rule out one or more candidate diagnoses of a differential diagnosis as a correct diagnosis for the patient.
11 . The computer-implemented method of claim 10 , the method further comprising:
providing, as additional input to the diagnosis model, results data of the one or more tests administered on the patient; and
generating, with the diagnosis model, another differential diagnosis including only a subset of the candidate diagnoses based on the results data.
12 . The computer-implemented method of claim 9 ,
wherein the input device includes one or more of:
a microphone that captures sounds generated by the patient;
a visual camera that captures images of the patient;
a thermometer that measures a body temperature of the patient;
a blood pressure meter that measures a blood pressure of the patient;
an oximeter that measures an oxygen concentration in blood of the patient; and
a pulse meter that measures a heart rate of the patient; and
wherein the patient intake data is elicited from the patient with a virtual assistant that converses with the patient and verbally asks the patient questions relating to the one or more physical symptoms and to a family medical history of the patient.
13 . The computer-implemented method of claim 9 , further comprising:
accessing, with the diagnosis model and from at least one database, a collection of at least 7,000 candidate human medical diagnoses.
14 . The computer-implemented method of claim 9 , further comprising:
determining that a third subset of the patient intake data relates to one or more medication side effects,
wherein at least one of the plurality of trained machine learning models compares the third subset of the patient intake data to medication data stored in a database, the medication data mapping medications to side effects of the medications and contraindications of the medications; and
wherein the diagnosis is generated also based on the third subset of patient intake data and a comparison of the third subset of patient intake data to the medication data.
15 . The computer-implemented method of claim 14 ,
wherein generation of the diagnosis includes mapping the third subset of data and the second subset of data to each of a plurality of candidate diagnoses.
16 . The computer-implemented method of claim 15 , wherein generation of the diagnosis includes simultaneous processing, by the plurality of machine learning models, of the first subset of data and the second subset of data.
17 . The computer-implemented method of claim 9 , further comprising:
outputting, by at least one of the plurality of trained machine learning models, and based on the first subset of the patient intake data and the second subset of the patient intake data, a recommendation to deprescribe a medication.