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

Paper Details

Abstracts, authors, affiliations, and disclosures.

Sunday, December 13, 2026
04:30 p.m. - 06:00 p.m.
MCCNO - R06-R09
Treatment Considerations, Supportive Care, Biological Therapies, Transplantation, Artificial Intelligence (AI), Machine Learning, Clinical Practice (Health Services and Quality), Health Economics, Quality Improvement, Quality of Care, Diseases, Myeloid Malignancies, Acute Myeloid Malignancies, Research, Clinical Research, Health Outcomes Research, Patient-Reported Outcomes, Study Population, Human
PAPER

737 Predicting survival and quality of life of older adults with acute myeloid leukemia (AML) using causal machine learning and patient-reported outcomes

Sara M. Tinsley-Vance, PhD, APRN, AOCN, Yi Luo, PhD2, Ram Thapa, PhD3, Cecile A Lengacher, RN, PhD, FAAN, FAPOS4, Steven K. Sutton, PhD5, Onyee Chan, MD6, Rami Komrokji, MD7, Eric Padron, MD8, David Sallman, MD9, Andrew T Kuykendall, MD10, Alison Walker, MD11, Kathryn Herricks, APRN12, Daniel Lastorino, BA13, Kristan Kersey, PA-C14, Heather Buffington, DNP15, Martine Extermann, MD, PhD16, Phuc Dinh Le17, Areej R El-Jawahri, MD18, Jeffrey Lancet, MD19
1 H. Lee Moffitt Cancer Center and Research Institute, Tampa, Florida, United States
2 H. Lee Moffitt Cancer Center and Research Institute, Tampa, Florida, United States
3 H. Lee Moffitt Cancer Center and Research Institute, Tampa, Florida, United States
4 University of South Florida, Tampa, Florida, United States
5 H. Lee Moffitt Cancer Center and Research Institute, Tampa, Florida, United States
6 H. Lee Moffitt Cancer Center and Research Institute, Tampa, Florida, United States
7 H. Lee Moffitt Cancer Center and Research Institute, Tampa, Florida, United States
8 H. Lee Moffitt Cancer Center and Research Institute, Tampa, Florida, United States
9 H. Lee Moffitt Cancer Center and Research Institute, Tampa, Florida, United States
10 H. Lee Moffitt Cancer Center and Research Institute, Tampa, Florida, United States
11 H. Lee Moffitt Cancer Center and Research Institute, Tampa, Florida, United States
12 H. Lee Moffitt Cancer Center and Research Institute, Tampa, Florida, United States
13 H. Lee Moffitt Cancer Center and Research Institute, Tampa, Florida, United States
14 H. Lee Moffitt Cancer Center and Research Institute, Tampa, Florida, United States
15 H. Lee Moffitt Cancer Center and Research Institute, Tampa, Florida, United States
16 H. Lee Moffitt Cancer Center and Research Institute, Tampa, Florida, United States
17 H. Lee Moffitt Cancer Center and Research Institute, Tampa, Florida, United States
18 Massachusetts General Hospital, Boston, Massachusetts, United States
19 H. Lee Moffitt Cancer Center and Research Institute, Tampa, Florida, United States

Background: Personalized treatment decision-making is a major unmet need in older adults with acute myeloid leukemia (AML), a population characterized by marked heterogeneity in disease characteristics, frailty, symptom burden, functional status, quality of life (QOL) and treatment preferences. Current prognostic models are predominantly focused on survival and treatment response. They fail to incorporate patient-reported outcomes (PROs) or QOL, despite their importance to patients when evaluating treatment options. 

Causal graphical models are uniquely suited to this challenge because they can integrate large, multidimensional datasets, robustly handle missing data, predict AML patients’ survival and QOL, identify causal-effect relationships among them, and determine personalized treatments to improve their therapeutical satisfaction based on individuals’ survival and QOL preference. While our long-term goal is to develop personalized treatment decision support for older adults with AML, this project is intended to develop a preliminary causal graphical model to predict individual AML patients’ survival and QOL.

Methods: As part of an NIH-funded prospective study of older adults with newly diagnosed AML, baseline clinical, disease-related, frailty assessment, and PRO data were collected before initiation of therapy, intensive compared to non-intensive. PRO measures included Brief Fatigue Inventory (BFI), Memorial Symptom Assessment Scale-Short Form (MSAS-SFI), and Functional Assessment of Cancer Therapy-Leukemia version (FACT-Leu Total). Consistent with the study's third aim, the baseline values of these variables were used to develop a causal Bayesian network (BN) as follows to predict the probabilities of individual AML patients’ two-year survival and QOL improvement simultaneously on Day 30. Here, the improvement of QOL is determined by whether the increment of the FACT-Leu value between Day 30 and the baseline is greater than 11 or not. The assumptions of our causal BN include non-unmeasured confounders and true representation of data-generating process.

First, we selected a set of causes most relevant to individual AML patient’s two-year survival or QOL improvement by integrating feature section approaches based on extended Markov blankets, Random Forests, and Least Absolute Shrinkage and Selection Operator (LASSO). We then identified the important factors for causal model building from the overlap of these two subsets based on expert knowledge.

Second, we employed a backward approach to learn the causal BN’s structure based on its prediction performance by eliminating relatively unimportant nodes in the network. 10-fold cross validation was used to evaluate the prediction performance of causal BN based on area under the free-response receiver operating characteristic (AU-FROC) curve.  

Third, we calibrated the causal BN based on robustness analysis and conditional independent tests to ensure that its implied conditional independencies match statistical distributions found from our collected data.

Results: The study included 171 patients (82 intensive-treatment, 89 non-intensive-treatment) with a mean age of 70.6±6.5 years. Patients receiving non-intensive therapy were significantly older than those receiving intensive therapy (74.0 vs. 67.0 years, p<0.001). The cohort was predominantly male (65.5%), White non-Hispanic (91.1%), married (71.6%), retired (77.1%), and highly educated, with 51.5% holding a bachelor's degree or higher. Baseline sociodemographic characteristics were otherwise similar between treatment groups.

In the causal BN model, factor “Therapy” is the direct cause of individual AML patients’ two outcomes, factors “MARITAL_Status” and “MSAS_PHYS” or factors “DISEASE3” and “BFI_Score” are the direct causes of the QOL or the survival. The rest of factors in the causal BN are their common indirect causes. While the AU-FROC value of two outcomes was 0.71, the prediction performance of survival and QOL were 0.67 and 0.63, respectively. Their performance can be further improved with additional PROs during the treatment.

Conclusions: We developed an accurate, explainable, and robust causal graphical model integrating AML patients’ clinical, disease-related, geriatric assessment, and PRO data. In addition to predicting their two-year survival and QOL improvements on Day 30, our model can be extended to support personalized treatment decision-making in older adults with AML before and during the treatment.
Sara M. Tinsley-Vance, PhD, APRN, AOCN: No relevant conflicts of interest to declare.