Systems and methods for improving retrieval- augmented generation in clinical decision support
Described are systems and methods for artificial intelligence (AI)-based clinical decision support. Systems can include a platform configured with a user input processing module, a context matching module, a retrieval-augmented generation (RAG) module, and an output generation module. Outputs of the platform can include a differential diagnosis, an assessment and treatment plan, or a clinical reference. The platform can further include an AI-copilot module and an AI-notebook module.
1. An artificial intelligence (AI)-based clinical decision support platform for treating a patient with a health condition, the platform comprising:
a user input processing module configured to receive a clinical problem representation (CPR) based at least in part on an input provided by a healthcare provider (HCP) or a patient, wherein the CPR comprises a list of one or more symptoms presented or determined in real time during a patient encounter;
a context matching module comprising a first large language model (LLM) generally trained using general or historical data, wherein the first LLM is configured to determine a clinical diagnosis of the health condition by (1) iteratively processing a first set of prompts using a plurality of healthcare context topics generated from a physician-validated context library and (2) determining, using the first set of prompts, a context match from the plurality of healthcare context topics, wherein the context match associates the health condition with at least one of the plurality of healthcare context topics with a confidence level;
a retrieval-augmented generation (RAG) module comprising the physician-validated context library, wherein the RAG module is configured to: (1) generate the plurality of healthcare context topics for use by the context matching module and (2) identify physician-validated schemas and clinical guidelines for treating the health condition; and
an output generation module comprising a second LLM trained using the physician-validated context library, wherein the second LLM is configured to determine a plurality of treatment plans for treating the health condition by (1) iteratively processing a second set of prompts using the context match and the physician-validated schemas and clinical guidelines and (2) determining, using the second set of prompts, at least one treatment plan of the plurality of treatment plans,
wherein the at least one treatment plan is used to treat the patient having or suspected of having the health condition based at least on a confidence level of the clinical diagnosis.
2. The platform of claim 1 , wherein the user input processing module is further configured to generate the CPR.
3. The platform of claim 1 , wherein the first set of prompts comprises a prompt for use by the first LLM to generate a differential diagnosis.
4. The platform of claim 1 , wherein the second set of prompts comprises a prompt for use by the second LLM to generate an assessment for the health condition.
5. The platform of claim 1 , wherein the RAG module is further configured to dynamically provide the plurality of healthcare context topics iteratively back to (i) the context matching module for subsequent use by the first LLM or (ii) the output generation module for subsequent use by the second LLM.
6. The platform of claim 1 , further comprising an AI-copilot module configured to generate data for a pre-visit clinical decision, wherein the data is received by the user input processing module to further assist the first LLM in generating the clinical diagnosis of the health condition or the second LLM in generating the at least one treatment plan for the health condition.
7. The platform of claim 1 , further comprising an AI-notebook module configured to generate data associated with clinical knowledge, wherein the data is received by the user input processing module to further assist the first LLM in generating the clinical diagnosis of the health condition or the second LLM in generating the at least one treatment plan for the health condition.
8. The platform of claim 1 , wherein the RAG module is further configured to use embedding-based matching to determine a relevance of the input provided by the HCP or the patient.
9. The platform of claim 1 , wherein the RAG module is further configured to use one or more prompts to improve association of the context match to the plurality of healthcare context topics.
10. The platform of claim 9 , wherein the first set of prompts, the second set of prompts, or the one or more prompts comprises chain-of-thought prompt engineering.
11. The platform of claim 1 , wherein the RAG module is further configured to use N-shot learning to improve association of the context match to the plurality of healthcare context topics.
12. The platform of claim 11 , wherein the N-shot learning comprises at least one learning example.
13. The platform of claim 1 , wherein the input provided by the HCP or the patient comprises a text input, a speech input, a video input, or any combination thereof.
14. The platform of claim 13 , wherein the platform is configured to (i) transform the speech input or the video input into a text input and (ii) transmit the text input to the user input processing module.
15. The platform of claim 5 , wherein a number of iterations is based at least on a confidence level of the clinical diagnosis of the health condition or the at least one treatment plan for treating the health condition.
16. The platform of claim 15 wherein the confidence level is at least 80%.
17. The platform of claim 1 , wherein the platform further comprises a graphical user interface (GUI) configured to display one or more structured healthcare reports to the HCP or the patient, and wherein the reports are structured into categories comprising the CPR, an assessment of the health condition, the clinical diagnosis of the health condition, and the at least one treatment plan for the health condition.
18. The platform of claim 1 , wherein the CPR further comprises age, medical history, medications, associated symptoms, lab tests, or a course of the health condition.
19. The platform of claim 1 , wherein the physician-validated context library is dynamically updated in real time for use by the context matching module.
20. A system comprising at least one processor and instructions executable by the at least one processor to cause the at least one processor to perform operations comprising:
(a) receiving a clinical problem representation (CPR) based at least in part on an input provided by a healthcare provider (HCP) or a patient, wherein the CPR comprises a list of one or more symptoms presented or determined in real time during a patient encounter;
(b) determining a clinical diagnosis by using the input to (1) iteratively process a first set of prompts using a plurality of healthcare context topics generated from a physician-validated context library and (2) determine, by a first large language model (LLM) using the first set of prompts, a context match from the plurality of healthcare context topics, wherein the context match associates the health condition with at least one of the plurality of healthcare context topics with a confidence level, and wherein the first LLM is generally trained using general or historical data;
(c) using retrieval augmented generation (RAG) to (1) generate the plurality of healthcare context topics for use in determining the clinical diagnosis in (b) and (2) identify physician-validated schemas and clinical guidelines for treating the health condition; and
(d) determining a plurality of treatment plans by using the physician-validated context library to (1) iteratively process a second set of prompts using the context match and the physician-validated schemas and clinical guidelines and (2) determine, by a second LLM using the second set of prompts, at least one treatment plan of the plurality of treatment plans, wherein the second LLM is trained using the physician-validated context library,
wherein the at least one treatment plan is used to treat the patient having or suspected of having the health condition based at least on a confidence level of the clinical diagnosis.
21. A computer-implemented method of providing artificial intelligence (AI)-based clinical decision support for treating a patient with a health condition, the method comprising:
(a) receiving a clinical problem representation (CPR) based at least in part on an input provided by a healthcare provider (HCP) or a patient, wherein the CPR comprises a list of one or more symptoms presented or determined in real time during a patient encounter;
(b) determining a clinical diagnosis by using the input to (1) iteratively process a first set of prompts using a plurality of healthcare context topics generated from a physician-validated context library and (2) determine, by a first large language model (LLM) using the first set of prompts, a context match from the plurality of healthcare context topics, wherein the context match associates the health condition with at least one of the plurality of healthcare context topics with a confidence level, and wherein the first LLM is generally trained using general or historical data;
(c) using retrieval augmented generation (RAG) to (1) generate the plurality of healthcare context topics for use in determining the clinical diagnosis in (b) and (2) identify physician-validated schemas and clinical guidelines for treating the health condition; and
(d) determining a plurality of treatment plans by using the physician-validated context library to (1) iteratively process a second set of prompts using the context match and the physician-validated schemas and clinical guidelines and (2) determine, by a second LLM using the second set of prompts, at least one treatment plan of the plurality of treatment plans, wherein the second LLM is trained using the physician-validated context library,
wherein the at least one treatment plan is used to treat the patient having or suspected of having the health condition based at least on a confidence level of the clinical diagnosis.