AI - Strategies: Personalization & Recommender Systems use cases from the ACM RecSys 2024 Health Recommender Systems workshop

The 6th Workshop on Health Recommender Systems, co-located with ACM RecSys in October 2024, showcased a diverse range of papers across several healthcare use cases which provide valuable insights for developing AI strategies for Personalization and Recommender Systems in healthcare.

Below are my mini-reviews of the papers:

1.        Personalized Health Recommender Systems:

Tailoring Health: Contextual Variables In Health Recommender Systems discusses the importance of data such as age, gender, weight, and ailments, and organizes them across two axes, objective-subjective and static-dynamic. A good framework to understand the impact of variables in recsys systems.

Improving the Prediction of Individual Engagement in Recommendations Using Cognitive Models proposes an Instance-based Learning (IBL) cognitive model which is created for every individual and better captures behavioral dynamics (vs LSTMs). These cognitive models aren’t frequently used in RecSys so might be worth exploring.

Enhancing Health Recommendations through Patient Metadata Integration: A Persona-Based Evaluation Approach shares a key insight that including trajectory or longitudinal data such as disease trajectory in user profiles plays a key role in creating dynamic, context-aware RecSys that evolve with a user’s health status.

2.        Content Recommendations:

Recommending News Articles for Public Health Intelligence uses event classifier labels (health policy, disease outbreaks, et al) on articles and graph clustering to recommend public health articles. The event classification seems to be the unique contribution here.

Health Document Presentation in Patient-Centered Recommender Systems with Carousel Interfaces - health article titles and summaries can be long and confusing. This paper shares an interesting approach using LLMs to generate AI-summaries of health articles to make them more presentable in carousel interfaces.

3.        Physical Activity and Exercise Recommendations:

Explaining Decision-Making between Exploration and Repetition: Key Factors for Physical Activity Recommendations analyzes factors that influence exploration (trying new activities) and repetition (sticking with familiar activities) for physical activities. For e.g., users were more likely to explore new activites on Sundays and that time related factors drove decisions for exploring new activities.

Prompting Large Language Models for Tailored Exercise Recommendations in Office Spaces identifies any posture anomalies from user videos and uses LLM for recommending exercises to address inadequate sitting postures. The combination of video posture analysis and LLM recommendations creates both complexities and opportunities in this use case.

Personalizing Exercise Recommendations with Explanations using Multi-Armed Contextual Bandit and Reinforcement Learning uses reinforcement learning with clinical guidelines to generate synthetic user profiles to address the cold start problem and has an explainable AI layer (xAI) to explain recommendations. The xAI layer is interesting but details are lacking in terms of how it was implemented.

4.        Nutrition and Food Recommendations:

The keynote Sustainable Food Recommendations Exploiting Natural Language Processing and Large Language Models provided an excellent overview of how to leverage a holistic user profile which includes food preferences, a knowledge base of food recipes, sustainability goals (e.g., carbon and water footprint), and an LLM for recipe selection to provide users with healthy food recipes.

Advancing Visual Food Attractiveness Predictions for Healthy Food Recommender Systems studies features of food images such as colorfulness and naturalness (vs mushy, slop, messy) which can nudge users towards healthy eating goals. Might be effective in multi-modal RecSys. Some food images can be mouth watering while others can be off-putting despite the actual flavor and taste.

5.        Clinical and Medical Recommendations:

Towards Recommender System Supported Contact Tracing for Cost-Efficient and Risk Aware Infection Suppression proposes Deep Reinforcement Learning to tackle the real-time decision-making challenges for contact tracing which modeled as a Markov decision process due to uncertainties inherent in the process. A promising public health use case that should be further fleshed out.

Enriching Clinical Sample Analysis with Biological Knowledge Graphs: A Preliminary Study uses biomedical knowledge graphs such as Reactome to understand clinical data and recommend proteins, pathways and interactions that researchers should investigate. A showcase for the power of knowledge graphs in clinical settings.

Design and Assessment of Representative Hybrid Clinical Trials using Health Recommender System uses real world data and a binary logistic regression based recsys for patient matching to ensure representativeness of the synthetic control groups and for equity adjustment. A use case which has a lot of potential for RecSys applications.