AI-Risk-Assistant

App Files

Application executable: Media: AI-Risk-Assistant.zip

Reference file: Media: openFDA device_search_fields.xlsx

Reference file: Media: openFDA recall_search_fields.xlsx

Settings.txt

CONFIG:OpenFDA-Vertex-V1.0 // identifies schema version (readonly)

https_url1: // openFDA api for MAUDE source data

https_url2: // openFDA api for Recall source data

limit1: // limit of MAUDE query results

limit2: // limit of Recall query results

report_title: // title of output report

data1_fields: // data fields to be analyzed in MDR query results

data2_fields: // Data fields analyzed in Recall query results

KEYWORD1-SEARCH: // general keyword search terms for MAUDE (matched in any field)

KEYWORD1-SEARCH: // general keyword search terms for Recall (matched in any field)

SEARCH1-FIELDS:SEARCH1-TERMS // MAUDE query search fields and terms, see openFDA device_search_fields.xlsx for full details of searchable fields

adverse_event_flag: // MAUDE example field

date_received: // MAUDE example field

device.device_report_product_code: // example field

SEARCH2-FIELDS:SEARCH2-TERMS // Recall query search fields and terms, see openFDA recall_search_fields.xlsx for full details of searchable fields

event_date_created: // Recall example field

recalling_firm: // Recall example field

reason_for_recall: // Recall example field

SORT1-FIELD:SORT1-TERM // MAUDE query results sorting date_received:desc

SORT2-FIELD:SORT2-TERM // Recall query results sorting event_date_created:desc

COUNT1-FIELD:COUNT1-TERM // MAUDE query count field/term

COUNT2-FIELD:COUNT2-TERM // Recall query count field/term

AI-ProblemSummaryPrompt: // Prompt that summarizes the product problem

AI-CountSummaryPrompt: // Prompt that counts the instances of items

AI-RecallSimilarityPrompt: // Prompt that analyzes the similarity between product problem and recalls

AI-WordsPerReport: // Number of words in each intermediate report (recommended: 1500, range: 100-5000)

AI-ModelTemperature: // LLM Temperature index (recommended: 0.05, range: 0..1f)

AI-ModelTopP: // LLM TOP_P index (recommended: 0.4, range: 0..1f)

AI-ModelTopK: // LLM TOP_K number of words for next word prediction (recommended: 10, range: 1-128)

AI-ModelMaxOutputTokens: // LLM maximum output words (range: 1-2048)

---

See the sample files page for example settings and reports.

Sample Files: AI-Risk-Assistant

App Description

App Use Case: The AI-Risk-Assistant collects and analyzes vast amounts of data from diverse sources to pinpoint safety trends and device issues, enabling proactive risk mitigation in premarket and post market risk management activities (e.g., PMS - post market surveillance). Utilizing AI capabilities, and access to abundant public data such as Recalls Enforcement and MAUDE data (courtesy of FDA), the Risk Assistant offers powerful tools to see the bigger picture and make better informed decisions. The Risk-Assistant easily consolidates key problems and potential harms associated with specific product codes, brands, or multiple devices under the same regulation, and presents those hidden insights and data patterns in an easy-to-understand format. This aids the risk management professional in balancing risks against benefits, and adopting a quantitative, data-driven approach, free from biases.

App Functionality: The AI-Risk-Assistant is designed to process product risk information, which may include details such as device models, usage settings, product codes/registrations, and other pertinent data. A Settings file contains the specific MAUDE and Recalls database query parameters as well as the AI prompts to process the data and identify hidden patterns and insights. With this information, the retrieval augmented generation model correlates the MAUDE and Recalls data sets and maps this information in a table format for easy consumption in risk management processes, such as hazard analyses, FMEAs (Failure Mode Effects Analysis) and other tools. The correlation is influenced by product codes, problem codes, specific malfunctions, use errors, root causes and reasons for recall, among other factors found in the data set and specified in the Settings file.

Diagram outlining the processing steps of the Risk Assistant:

An Excel document is generated containing results detailing the top product problems, along with additional analytics for each product problem including correlated produce recalls. It also includes the correlation information and hidden patterns in the MAUDE and Recalls data.