Mohan Purushothaman

Background/Objectives: In clinical informatics, the term ‘information overload’ is increasingly used to describe the operational impediments of excessive documentation. While electronic health records (EHRs) are growing in abundance, many medical records (MRs) remain in legacy formats that impede efficient, systematic processing, contributing to the extenuating challenges of care fragmentation. Thus, there is a growing interest in using generative AI (genAI) for automated MR summarization and characterization. Methods: MRs for a set of 78 individuals were digitized. Some were known systemic lupus erythematosus (SLE) cases, while others were under evaluation for possible SLE classification. A two-pass genAI assessment strategy was implemented using the Claude 3.5 large language model (LLM) to mine MRs for information relevant to classifying SLE vs. undifferentiated connective tissue disorder (UCTD) vs. neither via the 22-criteria EULAR 2019 model. Results: Compared to clinical determination, the antinuclear antibody (ANA) criterion (whose results are crucial for classifying SLE-negative cases) exhibited favorable sensitivity 0.78 ± 0.09 (95% confidence interval) and a positive predictive value 0.85 ± 0.08 but a marginal performance for specificity 0.60 ± 0.11 and uncertain predictivity for the negative predictive value 0.48 ± 0.11. Averaged over the remaining 21 criteria, these four performance metrics were 0.69 ± 0.11, 0.87 ± 0.04, 0.54 ± 0.10, and 0.93 ± 0.03. Conclusions: ANA performance statistics imply that genAI yields confident assessments of SLE negativity (per high sensitivity) but weaker positivity. The remaining genAI criterial determinations support (per specificity) confident assertions of SLE-positivity but tend to misclassify a significant fraction of clinical positives as UCTD.

Systemic lupus erythematosus (SLE) is a complex autoimmune disease that poses serious long-term patient burdens. (1) Background: SLE patient classification and care are often complicated by case heterogeneity (diverse variations in symptoms and severity). Large language models (LLMs) and generative artificial intelligence (genAI) may mitigate this challenge by profiling medical records to assess key medical criteria. (2) Methods: To demonstrate genAI-based profiling, ACR (American College of Rheumatology) 1997 SLE classification criteria were used to define medically relevant LLM prompts. Records from 78 previously studied patients (45 classified as having SLE; 33 indeterminate or negative) were computationally profiled, via five genAI replicate runs. (3) Results: GenAI determinations of the “Discoid Rash” and “Pleuritis or Pericarditis” classification criteria yielded perfect concurrence with clinical classification, while some factors such as “Immunologic Disorder” (56% accuracy) were statistically unreliable. Compared to clinical classification, our genAI approach achieved a 72% predictive success rate. (4) Conclusions: GenAI classifications may prove sufficiently predictive to aid medical professionals in evaluating SLE patients and structuring care strategies. For individual criteria, accuracy seems to correlate inversely with complexities in clinical determination, implying that improvements in AI patient profiling tools may emerge from continued advances in clinical classification efficacy.

(1) Objective: Systemic lupus erythematosus (SLE) is a complex disease involving immune dysregulation, episodic flares, and poor quality of life (QOL). For a decentralized digital study of SLE patients, machine learning was used to assess patient-reported outcomes (PROs), QOL, and biometric data for predicting possible disease flares. (2) Methods: Participants were recruited from the LupusCorner online community. Adults self-reporting an SLE diagnosis were consented and given a mobile application to record patient profile (PP), PRO, and QOL metrics, and enlisted participants received smartwatches for digital biometric monitoring. The resulting data were profiled using feature selection and classification algorithms. (3) Results: 550 participants completed digital surveys, 144 (26%) agreed to wear smartwatches, and medical records (MRs) were obtained for 68. Mining of PP, PRO, QOL, and biometric data yielded a 26-feature model for classifying participants according to MR-identified disease flare risk. ROC curves significantly distinguished true from false positives (ten-fold cross-validation: p < 0.00023; five-fold: p < 0.00022). A 25-feature Bayesian model enabled time-variant prediction of participant-reported possible flares (P(true) > 0.85, p < 0.001; P(nonflare) > 0.83, p < 0.0001). (4) Conclusions: Regular profiling of patient well-being and biometric activity may support proactive screening for circumstances warranting clinical assessment.