LLMD: A Large Language Model for Interpreting Longitudinal Medical Records

We introduce LLMD, a large language model designed to analyze a patient's medical history based on their medical records. Along with domain knowledge, LLMD is trained on a large corpus of records collected over time and across facilities, as well as tasks and labels that make nuanced connections among them. This approach is critical to an accurate picture of patient health, and has distinctive advantages over models trained on knowledge alone, unlabeled records, structured EHR data, or records from a single health system. The recipe for LLMD continues pretraining a foundational model on both domain knowledge and the contents of millions of records. These span an average of 10 years of care and as many as 140 care sites per patient. LLMD is then instruction fine-tuned on structuring and abstraction tasks. The former jointly identify and normalize document metadata, provenance information, clinical named-entities, and ontology mappings, while the latter roll these into higher-level representations, such a continuous era of time a patient was on a medication. LLMD is deployed within a layered validation system that includes continual random audits and review by experts, e.g. based on uncertainty, disease-specific rules, or use-case. LLMD exhibits large gains over both more-powerful generalized models and domain-specific models. On medical knowledge benchmarks, LLMD-8B achieves state of the art accuracy on PubMedQA text responses, besting orders-of-magnitude larger models. On production tasks, we show that LLMD significantly outperforms all other models evaluated, and among alternatives, large general purpose LLMs like GPT-4o are more accurate than models emphasizing medical knowledge. We find strong evidence that accuracy on today's medical benchmarks is not the most significant factor when analyzing real-world patient data, an insight with implications for future medical LLMs.'

Evaluating Fairness Metrics in the Presence of Dataset Bias

Data-driven algorithms play a large role in decision making across a variety of industries. Increasingly, these algorithms are being used to make decisions that have significant ramifications for people's social and economic well-being, e.g. in sentencing, loan approval, and policing. Amid the proliferation of such systems there is a growing concern about their potential discriminatory impact. In particular, machine learning systems which are trained on biased data have the potential to learn and perpetuate those biases. A central challenge for practitioners is thus to determine whether their models display discriminatory bias. Here we present a case study in which we frame the issue of bias detection as a causal inference problem with observational data. We enumerate two main causes of bias, sampling bias and label bias, and we investigate the abilities of six different fairness metrics to detect each bias type. Based on these investigations, we propose a set of best practice guidelines to select the fairness metric that is most likely to detect bias if it is present. Additionally, we aim to identify the conditions in which certain fairness metrics may fail to detect bias and instead give practitioners a false belief that their biased model is making fair decisions.