2026 ASH®
Annual Meeting & Exposition December 12-15, 2026 ▪ New Orleans, Louisiana

Paper Details

Abstracts, authors, affiliations, and disclosures.

Sunday, December 13, 2026
06:00 p.m. - 08:00 p.m.
MCCNO - Hall F
Access, Representation, Population Health, Technology and Procedures, Serologic Tests, Blood Banking, Treatment Considerations, Supportive Care, Biological Therapies, Transfusion, Gene Therapy, Non-Biological Therapies, Opioids, Workforce, Artificial Intelligence (AI), Machine Learning, Clinical Practice (Health Services and Quality), Clinical Guidelines, Health Economics, Patient Safety, Practice Models, Quality Improvement, Quality of Care, Systems-Based Hematology, Telemedicine, Diseases, Hemoglobinopathies, Sickle Cell Disease, Metabolic Disorders, Iron Overload, Research, Clinical Research, Health Outcomes Research, Patient-Reported Outcomes, Real-World Evidence, Registries, Study Population, Human, Adult, Pediatric
PAPER

4358 The development and end-user testing of spright 2.0, a machine learning clinical decision support tool for informed shared decision-making for hematopoietic cell transplantation for sickle cell disease

Shivapriya Somasundaram, MS, Rajagopal Subramaniam Chandrashekhar, M. Tech2, Lakshmanan Krishnamurti, Professor Pediatrics,3
1 Yale School of Medicine, New Haven, Connecticut, United States
2 Yale School of Medicine, New Haven, Connecticut, United States
3 Yale School of Medicine, New Haven, Connecticut, United States

 OBJECTIVES: Hematopoietic cell transplantation is a treatment with curative intent for sickle cell disease (SCD). However, complex risk-benefit tradeoffs, variable personal preferences, and the heterogeneity of outcomes all contribute to decisional dilemmas. We therefore developed SPRIGHT, a machine learning (ML) predictive model with high predictive performance (Chandrasekhar et al JMIR AI 2025). We undertook to translate this predictive tool into a workflow-integrated, interpretable, preference-sensitive guidance that clinicians and patients can use to make an informed shared decision and describe the development and end-user testing of the CDS tool; SPRIGHT 2.0.

METHODS: We trained the prediction model on the CIBMTR HCT for SCD dataset, repeated the internal cross-validation, and implemented it in Gradio, a lightweight open-source Python library for creation of shareable web functions for testing ML models. We hosted the application on Hugging Face (Hugging Face Inc., NY, NY), a platform designed for collaboration for ML models and applications. We developed the following features in creating the SPRIGHT 2.0 CDS: 1. Individualized prediction of outcomes with confidence intervals, visualized with icon arrays for patient explanation, and SHAP plots to describe feature importance. 2. Counterfactual calculations and display of likely outcomes when different treatment strategies are applied for a patient with given baseline characteristics.3. A large language model (LLaMA 3.3 70B Meta AI 2023) generated, downloadable, plain language patient directed summary, physician directed summary and recommendations, predicted probabilities, counterfactual summaries. 4 Conversational chatbot text interface that allows users to query the HCT for SCD dataset using plain language queries which are converted into structured queries executed in python to query the dataset. This provides clinicians with an accessible interface for investigating patient populations and understanding the data underlying the CDS predictions. We conducted a mixed-methods evaluation with 11 clinical experts in HCT for SCD (median 14.6 years of experience) using iterative beta testing and semi-structured cognitive interviews. Each clinical expert completed the System Causability Scale (SUS), System Usability Scale (SCS), Trust in Automated Systems Scale (TIAS), and NASA-TLX. All Interviews were recorded, transcribed, coded, and subjected to qualitative thematic analysis using NVivo (Lumivero.com).

RESULTS: Participants reported high usability (SUS; 0-100, mean score of 80.22) and excellent internal consistency (Cronbach’s α = 0.9322). The quality of explainability was rated high (SCS scaled 0-1; mean = 0.77), with strong reliability (α = 0.874). Technology acceptance and trust were favorable, with the TIAS (scaled1-7, mean= 5.586)  and good internal consistency (α=0.8569). Workload  was assessed with the NASA-TLX (rated 1-21 and converted to a 0-100 metric, mean=17.76). Subscale means were Mental Demand (21.25), Temporal Demand (12.85), Effort (20.71), and Frustration (16.25). Emergent themes in qualitative analysis were: 1. Utility, with experts rating the CDS as highly clinically useful with strong potential to support shared decision-making. 2. Side-by-side counterfactual comparisons, which enable comparison of various transplant strategies given baseline donor characteristics, were considered a particularly useful feature. 3. Usability was high, with the interface found to be user-friendly with straightforward data entry and good workflow fit. 4. In analyzing the model, participants emphasized the need for clear, lay-friendly language and careful framing to avoid overstating model scope 5. Recommended enhancements include additional comparators (gene therapy, non-transplant approaches); additional baseline inputs and outcomes; time-to-event displays; greater data transparency; continual model correction and updates; and patient-facing governance and safety. We made iterative upgrades and continue to develop the CDS tool.

CONCLUSIONS: SPRIGHT 2.0 is a secure, web-based ML CDS platform that supports individualized, informed, shared decision-making in HCT for SCD with outcome predictions and integrates explainability and patient-oriented summaries in a unified interface, with data- and preference-driven recommendations in plain text. Future work will evaluate the impact of CDS in patient decision-making and the feasibility of dissemination and implementation.

Shivapriya Somasundaram, MS: No relevant conflicts of interest to declare.