MedCGA - Comprehensive Geriatric Assessment NLP Framework

We developed MedCGA, a multi-method NLP framework to identify Comprehensive Geriatric Assessment (CGA) and Age-Friendly Health Systems (AFHS) 4Ms data elements from unstructured clinical notes.

Definition of CGA and AFHS 4Ms Data Elements

Definition of CGA and AFHS 4Ms Tasks

Domain	Sub-task	Definition
GOC and Advanced Care Planning	goal_of_care	Discussions about goals of care, advance care planning, or patient values (What Matters).
	limitation_of_lst	Evidence of limitations to life-sustaining treatment (e.g., DNR/DNI status).
	hospice	Evidence of hospice enrollment, discussion, or eligibility.
	palliative_care	Discussion or involvement of specialist palliative care services.
Social-Economic Status	financial_strain	Financial hardship affecting ability to afford care or basic needs.
	food_insecurity	Limited or uncertain access to adequate and nutritious food.
	housing_insecurity	Unstable housing, risk of eviction, or inability to afford housing.
	social_family_support	Limited social or family support, including emotional or instrumental support.
	transportation_barrier	Difficulty accessing care due to transportation limitations.
Cognitive and Behavioral Status	delirium	Acute and fluctuating disturbance in mental status.
	agitation	Increased psychomotor activity with restlessness or aggressive behavior.
	amc	Altered mental status (non-specific cognitive or awareness change).
	confusion	Reduced clarity of thought, attention, or decision-making.
	disconnected	Disorganized or detached thought or perception.
	disorganized_thinking	Illogical or incoherent thought processes or speech.
	disorient	Impaired awareness of time, place, or identity.
	encephalopathy	Brain dysfunction causing altered mental state.
	fluctuation	Variation in symptom severity over short periods.
	hallucination	Perception without external stimulus.
	inattention	Reduced ability to focus or sustain attention.
	reoriented	Restoration of orientation after confusion or altered state.
	cognitive_impairment	Deficits in memory, attention, or executive function.
	depression	Persistent low mood with associated cognitive and physical symptoms.
Mobility and Functional Status	physical_activity	Engagement in bodily movement contributing to function and health.
	falls	Unintentional descent to the ground with or without injury.
	basic_adl	General impairment in activities of daily living.
	bathing	Difficulty washing or drying the body.
	dressing	Difficulty putting on or removing clothing.
	transferring	Difficulty changing positions or moving between surfaces.
	continence	Difficulty controlling bladder or bowel function.
	toileting	Difficulty using the toilet and managing hygiene.
	feeding	Difficulty eating or handling food.
	telephone_use	Difficulty using communication devices.
	using_transportation	Difficulty accessing or navigating transportation.
	shopping	Difficulty obtaining goods and services.
	preparing_food	Difficulty planning and preparing meals.
	housekeeping	Difficulty maintaining a clean and safe living environment.
	doing_laundry	Difficulty washing and managing clothes.
	handyman_work	Difficulty performing home maintenance tasks.
	managing_medication	Difficulty managing medication regimen.
	handling_finance	Difficulty managing financial tasks and transactions.

Methods

We implemented three approaches for CGA extraction:

1. Generative LLM (GPT-4o)

   • In-context learning
   • High sensitivity for complex and context-rich concepts

2. MedAgingIE (Symbolic + LLM Hybrid)

   • Ontology-guided rules
   • LLM-assisted rule construction
   • Balanced precision and recall

3. Instruction-Tuned Model (Qwen2-7B-Instruct)

   • Open-source model
   • Scalable but lower performance

Getting Started (MedAgingIE)

Prerequisites

Java 1.8

NLP framework MedTaggerIE

MedTagger contains a suite of programs that the Mayo Clinic NLP program has developed in 2013. It includes three major components: MedTagger for indexing based on dictionaries, MedTaggerIE for information extraction based on patterns, and MedTaggerML for machine learning-based named entity recognition.

Clone

https://github.com/OHNLP/MedCGA

Configuration

A step by step configuration instructions can be accessed through: https://vimeo.com/392331446

1. Move the .jar file to the Delirium folder
2. Modify the INPUTDIR, OUTPUTDIR, and RULEDIR variables in runMedTagger-fit-delirium.sh or runMedTagger-fit-delirium.bat, as appropriate
   • INPUT_DIR: full directory path of input folder
   • OUTPUT_DIR: full directory path of output folder
   • RULES_DIR: full directory path of 'Rule' folder

Run

runMedTagger-fit-MedCGA.sh

Refine

Due to the institutional-specific heterogeneity, after deplying the algorithm, we recommend 1) conducting local evaluation through manual chart review, 2) refining the keywords, 3) implementing a section detection algorithm based on the structure of clinical notes. Additional information the NLP deployment and evaluation process can be found at: https://github.com/OHNLP/annotation-best-practices

Reference

Fu S, et al. A Multi-site Benchmarking Framework for Scalable Extraction of Geriatric Care Constructs from Electronic Health Records