Use case: Managing and upskilling your healthcare staff

Implementing talent-transformation and upskilling strategies helps workforces remain adept at using new technologies and practices in medical and healthcare services. Proactive upskilling initiatives make sure that healthcare professionals can provide high-quality patient care, optimize operational efficiencies, and stay compliant with regulatory standards. Moreover, talent transformation fosters a culture of continuous learning. This is pivotal for adapting to the changing healthcare landscape and addressing emerging public health challenges. Traditional training approaches, such as classroom-based training and static learning modules, offer uniform content to a broad audience. They often lack personalized learning paths, which are critical for addressing the specific needs and proficiency levels of individual practitioners. This one-size-fits-all strategy can result in disengagement and suboptimal knowledge retention.

Consequently, healthcare organizations must embrace innovative, scalable, and technology-driven solutions that can determine the gap for each of their employees in their current state and potential future state. These solutions should recommend hyper-personalized learning pathways and the right set of learning content. This effectively prepares the workforce for the future of healthcare.

In the healthcare industry, you can apply generative AI to help you understand and upskill your workforce. Through the connection of large language models (LLMs) and advanced retrievers, organizations can understand what skills they currently have and identify key skills that might be necessary in the future. This information helps you bridge the gap by hiring new workers and upskilling the current workforce. Using Amazon Bedrock and knowledge graphs, healthcare organizations can develop domain-specific applications that facilitate continuous learning and skill development.

The knowledge provided by this solution helps you effectively manage talent, optimize workforce performance, drive organizational success, identify existing skills, and craft a talent strategy. This solution can help you perform these tasks in weeks instead of months.

Solution overview

This solution is a healthcare talent transformation framework that consists of the following components:

• Intelligent resume parser – This component can read a candidate's resume and precisely extract candidate information, including skills. Intelligent information extraction solution built using fine-tuned Llama 2 model in Amazon Bedrock on a proprietary training dataset covering resumes and talent profiles from more than 19 industries. This LLM-based process saves hundreds of hours by automating the manual review process of resumes and matching top candidates to open roles.

• Knowledge graph – A knowledge graph built on Amazon Neptune, a unified repository of talent information including role and skill taxonomy of the organization as well as the industry, capturing the semantics of healthcare talent using definitions of skills, roles and their properties, relations, and logical constraints.

• Skill ontology – The discovery of skill proximities between candidate skills and ideal current state or future state skills (retrieved using a knowledge graph) is achieved through ontology algorithms that measure the semantic similarity between candidate skills and target state skills.

• Learning pathway and content – This component is a learning recommendation engine that can recommend the right learning content from a catalogue of learning materials from any vendor based on the identified skill gaps. Identifying the most optimal upskilling pathways for each candidate by analyzing the skill gaps and recommending prioritized learning content, to enable a seamless and continuous professional development for each candidate during the transition to a new role.

This cloud-based, automated solution is powered by machine learning services, LLMs, knowledge graphs, and Retrieval Augmented Generation (RAG). It can scale to process tens or thousands of resumes in minimum amount of time, create instant candidate profiles, identify gaps in their current or potential future state, and then efficiently recommend the right learning content to close these gaps.

The following image shows the end-to-end flow of the framework. The solution is built on fine-tuned LLMs in Amazon Bedrock. These LLMs retrieve data from the healthcare talent knowledge base in Amazon Neptune. Data-driven algorithms make recommendations for optimal learning pathways for each candidate.

Building this solution consists of the following steps:

• Step 1: Extracting talent information and building a skills profile

• Step 2: Discovering role-to-skill relevance from a knowledge graph

• Step 3: Identifying skill gaps and recommending training

Step 1: Extracting talent information and building a skills profile

First, you fine-tune a large language model, such as Llama 2, in Amazon Bedrock with a custom dataset. This adapts the LLM for the use case. During training, you accurately and consistently extract key talent attributes from candidate resumes or similar talent profiles. These talent attributes include skills, current role title, experience titles with date spans, education, and certifications. For more information, see Customize your model to improve its performance for your use case in the Amazon Bedrock documentation.

The following image shows the process to fine-tune a resume-parsing model by using Amazon Bedrock. Both real and synthetically created resumes are passed to an LLM in order to extract key information. A group of data scientists validate the extracted information against the original, raw text. The extracted information is then concatenated by using chain-of-thought prompting and the original text to derive a training dataset for fine-tuning. This dataset is then passed to an Amazon Bedrock customization job, which fine-tunes the model. An Amazon SageMaker AI batch job runs a model-evaluation framework that evaluates the fine-tuned model. If the model needs improvement, the job runs again with more data or different hyperparameters. After the evaluation meets the standards, you host the custom Model through Amazon Bedrock provisioned throughput.

Step 2: Discovering role-to-skill relevance from a knowledge graph

Next, you build a knowledge graph that encapsulates the skills and role taxonomy of your organization and of other organizations in the healthcare industry. This enriched knowledge base is sourced from aggregated talent and organization data in Amazon Redshift. You can gather talent data from a range of labor-market data providers and from organization-specific structured and unstructured data sources, such as enterprise resource planning (ERP) systems, a human resources information system (HRIS), employee resumes, job descriptions, and talent architecture documents.

Build the knowledge graph on Amazon Neptune. Nodes represent skills and roles, and edges represent the relationships between them. Enrich this graph with metadata to include details such as organization name, industry, job family, skill type, role type, and industry tags.

Next, you develop a Graph Retrieval Augmented Generation (Graph RAG) application. Graph RAG is a RAG approach that retrieves data from a graph database. The following are the components of the Graph RAG application:

• Integration with an LLM in Amazon Bedrock – The application uses an LLM in Amazon Bedrock for natural language understanding and query generation. Users can interact with the system by using natural language. This makes it accessible to non-technical stakeholders.

• Orchestration and information retrieval – Use LlamaIndex or LangChain orchestrators to facilitate the integration between the LLM and the Neptune knowledge graph. They manage the process of converting natural language queries into openCypher queries. Then, they run the queries on the knowledge graph. Use prompt engineering to instruct the LLM about best practices for constructing openCypher queries. This helps optimize the queries to retrieve the relevant subgraph, which contains all of the pertinent entities and relationships about the queried roles and skills.

• Insight generation – The LLM in Amazon Bedrock processes the retrieved graph data. It generates detailed insights about the current state and projects future states for the queried role and associated skills.

The following image shows the steps to build a knowledge graph from source data. You pass the structured and unstructured source data to the data ingestion pipeline. The pipeline extracts and transforms information to a CSV bulk load formation that is compatible with Amazon Neptune. The bulk loader API uploads the CSV files that are stored in an Amazon S3 bucket to the Neptune knowledge graph. For user queries related to talent future state, relevant roles, or skills, the fine-tuned LLM in Amazon Bedrock interacts with the knowledge graph through a LangChain orchestrator. The orchestrator retrieves the relevant context from the knowledge graph and push the responses to the insights table in Amazon Redshift. The LangChain orchestrator, like GraphQAChain, converts the natural language query from the user to an openCypher query in order to query the knowledge graph. The Amazon Bedrock fine-tuned model generates a response based on the retrieved context.

Step 3: Identifying skill gaps and recommending training

In this step, you accurately compute the proximity between the current state of a healthcare professional and potential future state roles. To do this, you perform a skill affinity analysis by comparing the individual's skills sets with the job role. In an Amazon OpenSearch Service vector database, you store skill taxonomy information and skill metadata, such as the skill description, skill type, and skill clusters. Use an Amazon Bedrock embedding model, such as Amazon Titan Text Embeddings models, to embed the identified key skill into vectors. Through a vector search, you retrieve the descriptions of current state skills and target state skills and perform an ontology analysis. The analysis provides proximity scores between the current and target state skill pairs. For each pair, you use the computed ontology scores to identify the gaps in skill affinities. Then, you recommend the optimal path for upskilling, which the candidate can consider during role transitions.

For each role, recommending the correct learning content for upskilling or reskilling involves a systematic approach that begins with creating a comprehensive catalog of learning content. This catalog, which you store in an Amazon Redshift database, aggregates content from various providers and includes metadata, such as the content duration, difficulty level, and learning mode. The next step is to extract the key skills offered by each piece of content and then map them to the individual skills required for the target role. You achieve this mapping by analyzing the coverage provided by the content through a skills proximity analysis. This analysis assesses how closely the skills taught by the content align with the desired skills for the role. The metadata plays a critical role in selecting the most appropriate content for each skill, making sure that learners receive tailored recommendations that suit their learning needs. Use LLMs in Amazon Bedrock to extract skills from the content metadata, perform feature engineering, and validate the content recommendations. This improves accuracy and relevance in the upskilling or reskilling process.

Alignment to the AWS Well-Architected Framework

The solution aligns with all six pillars of the AWS Well-Architected Framework:

• Operational excellence – A modular, automated pipeline enhances operational excellence. Key components of the pipeline are decoupled and automated, allowing for faster model updates and easier monitoring. Additionally, automated training pipelines support quicker releases of fine-tuned models.

• Security – This solution processes sensitive and personally identifiable information (PII), such as the data in resumes and talent profiles. In AWS Identity and Access Management (IAM), implement fine-grained access control policies and make sure that only authorized personnel have access to this data.

• Reliability – The solution uses AWS services, such as Neptune, Amazon Bedrock, and OpenSearch Service, that provide fault tolerance, high availability, and uninterrupted access to insights even during high demand.

• Performance efficiency – Fine-tuned LLMs in Amazon Bedrock and OpenSearch Service vector databases are designed to quickly and accurately process large datasets in order to deliver timely, personalized learning recommendations.

• Cost optimization – This solution uses a RAG approach, which reduces the need for continuous pre-training of models. Instead of fine-tuning the entire model repeatedly, the system fine-tunes only specific processes, such as extracting information from resumes and structuring outputs. This results in significant cost savings. By minimizing the frequency and scale of resource-intensive model training and by using pay-per-use cloud services, healthcare organizations can optimize their operational costs while maintaining high performance.

• Sustainability – This solution uses scalable, cloud-native services that allocate compute resources dynamically. This reduces energy consumption and the environmental impact while still supporting large-scale, data-intensive talent transformation initiatives.