Heather Mattie

James Wen, Sahil Nalawade, Zhiwei Liang et al. · deepmind, harvard

Background: Generative artificial intelligence (AI) deployment in academic medical settings raises copyright compliance concerns. Dana-Farber Cancer Institute implemented GPT4DFCI, an internal generative AI tool utilizing OpenAI models, that is approved for enterprise use in research and operations. Given (1) the exceptionally broad adoption of the tool in our organization, (2) our research mission, and (3) the shared responsibility model required to benefit from Customer Copyright Commitment in Azure OpenAI Service products, we deemed rigorous copyright compliance testing necessary. Case Description: We conducted a structured red teaming exercise in Nov. 2024, with 42 participants from academic, industry, and government institutions. Four teams attempted to extract copyrighted content from GPT4DFCI across four domains: literary works, news articles, scientific publications, and access-restricted clinical notes. Teams successfully extracted verbatim book dedications and near-exact passages through various strategies. News article extraction failed despite jailbreak attempts. Scientific article reproduction yielded only high-level summaries. Clinical note testing revealed appropriate privacy safeguards. Discussion: The successful extraction of literary content indicates potential copyrighted material presence in training data, necessitating inference-time filtering. Differential success rates across content types suggest varying protective mechanisms. The event led to implementation of a copyright-specific meta-prompt in GPT4DFCI; this mitigation has been in production since Jan. 2025. Conclusion: Systematic red teaming revealed specific vulnerabilities in generative AI copyright compliance, leading to concrete mitigation strategies. Academic medical institutions deploying generative AI should implement continuous testing protocols to ensure legal and ethical compliance.

Heather Mattie, Patrick Reidy, Patrik Bachtiger et al.

It is not clear how to target patients who are most likely to benefit from digital care management programs ex-ante, a shortcoming of current risk score based approaches. This study focuses on defining impactability by identifying those patients most likely to benefit from technology enabled care management, delivered through a digital health platform, including a mobile app and clinician web dashboard. Anonymized insurance claims data were used from a commercially insured population across several U.S. states and combined with inferred sociodemographic data and data derived from the patient-held mobile application itself. Our approach involves the creation of two models and the comparative analysis of the methodologies and performances therein. We first train a cost prediction model to calculate the differences in predicted (without intervention) versus actual (with onboarding onto digital health platform) healthcare expenditure for patients (N = 1,242). This enables the classification of impactability if differences in predicted versus actual costs meet a predetermined threshold. A random forest machine learning model was then trained to accurately categorize new patients as impactable versus not impactable, reaching an overall accuracy of 71.9%. We then modify these parameters through grid search to define the parameters that deliver optimal performance. A roadmap is proposed to iteratively improve the performance of the model. As the number of newly onboarded patients and length of use continues to increase, the accuracy of predicting impactability will improve commensurately as more advanced machine learning techniques such as deep learning become relevant. This approach is generalizable to analyzing the impactability of any intervention and is a key component of realising closed loop feedback systems for continuous improvement in healthcare.

Heather Mattie, Kenth Engø-Monsen, Rich Ling et al.

Understanding tie strength in social networks, and the factors that influence it, have received much attention in a myriad of disciplines for decades. Several models incorporating indicators of tie strength have been proposed and used to quantify relationships in social networks, and a standard set of structural network metrics have been applied to predominantly online social media sites to predict tie strength. Here, we introduce the concept of the "social bow tie" framework, a small subgraph of the network that consists of a collection of nodes and ties that surround a tie of interest, forming a topological structure that resembles a bow tie. We also define several intuitive and interpretable metrics that quantify properties of the bow tie. We use random forests and regression models to predict categorical and continuous measures of tie strength from different properties of the bow tie, including nodal attributes. We also investigate what aspects of the bow tie are most predictive of tie strength in two distinct social networks: a collection of 75 rural villages in India and a nationwide call network of European mobile phone users. Our results indicate several of the bow tie metrics are highly predictive of tie strength, and we find the more the social circles of two individuals overlap, the stronger their tie, consistent with previous findings. However, we also find that the more tightly-knit their non-overlapping social circles, the weaker the tie. This new finding complements our current understanding of what drives the strength of ties in social networks.