DIRECT MEDICAL COSTS SAVED BY EARLY PAH DETECTION USING A MACHINE LEARNING ALGORITHM
Author(s)
Tanvee Mahesh Thakur, PhD1, Neto Coulibaly, MS1, Praveen Kumar-M, MD2, Krishnappa Purushotham, MBBS2, Sanjeev Shrinivas Nadapurohit, MBBS2, Jason Maron, BS, MBA, MPH, MS2, Anna Watzker, MHS1, Anna Hemnes, MD3, Hilary Dubrock, MD4.
1Merck & Co., Inc, Rahway, NJ, USA, 2nference, Cambridge, MA, USA, 3Vanderbilt University Medical Center, Nashville, TN, USA, 4Mayo Clinic, Rochester, MN, USA.
1Merck & Co., Inc, Rahway, NJ, USA, 2nference, Cambridge, MA, USA, 3Vanderbilt University Medical Center, Nashville, TN, USA, 4Mayo Clinic, Rochester, MN, USA.
OBJECTIVES: Pulmonary Arterial Hypertension (PAH) is a rare disease characterized by non-specific symptoms that often lead to delayed diagnosis and increased healthcare resource utilization. This study evaluates the pre-diagnosis direct medical costs (DMC) that could be saved from earlier PAH detection by a machine learning (ML) algorithm.
METHODS: Data from Mayo Clinic Electronic Health Record (2015-2024) were used in this retrospective study to identify adult PAH patients with at least two PAH ICD codes or clinical- note mentions of PAH using natural language processing (NLP) separated by 30 days, one PAH-specific medication, and one PAH-related symptom. The index date was the earliest ICD-coded or NLP-identified PAH diagnosis date. A CatBoost ML algorithm was used to detect PAH within one year of the earliest PAH symptom prior to clinical diagnosis, enabling DMC savings due to earlier detection. The key outcome was DMC saved Per Patient Per Month (PPPM), overall and by care setting, measured 12-months pre-index, and defined as the difference between ‘earliest symptom-to-diagnosis’ and ‘earliest symptom-to-algorithm detection’ costs.
RESULTS: Among 799 PAH patients, mean age (SD) was 62.7 (14.4), 59% were female, mean time from first symptom to algorithm detection was 0.6 (2.2) months, and from first symptom to clinical diagnosis was 4.0 (4.3) months. Overall, DMC saved due to earlier detection were (mean) $7,658 PPPM, driven primarily by in-patient, emergency department, and critical care visits ($4,455); followed by diagnostic imaging procedures ($1,022); and outpatient, office and remote visits ($273). Compared to patients with lowest DMC savings (bottom-quartile), those with highest DMC savings (top-quartile) were detected 3.7-months (mean) earlier, corresponding to approximately $13,540 PPPM greater DMC saved.
CONCLUSIONS: Earlier detection of PAH facilitated by this ML algorithm was associated with meaningful savings in DMC, indicating that such tools, when implemented within health systems, may support downstream resource allocation and cost savings.
METHODS: Data from Mayo Clinic Electronic Health Record (2015-2024) were used in this retrospective study to identify adult PAH patients with at least two PAH ICD codes or clinical- note mentions of PAH using natural language processing (NLP) separated by 30 days, one PAH-specific medication, and one PAH-related symptom. The index date was the earliest ICD-coded or NLP-identified PAH diagnosis date. A CatBoost ML algorithm was used to detect PAH within one year of the earliest PAH symptom prior to clinical diagnosis, enabling DMC savings due to earlier detection. The key outcome was DMC saved Per Patient Per Month (PPPM), overall and by care setting, measured 12-months pre-index, and defined as the difference between ‘earliest symptom-to-diagnosis’ and ‘earliest symptom-to-algorithm detection’ costs.
RESULTS: Among 799 PAH patients, mean age (SD) was 62.7 (14.4), 59% were female, mean time from first symptom to algorithm detection was 0.6 (2.2) months, and from first symptom to clinical diagnosis was 4.0 (4.3) months. Overall, DMC saved due to earlier detection were (mean) $7,658 PPPM, driven primarily by in-patient, emergency department, and critical care visits ($4,455); followed by diagnostic imaging procedures ($1,022); and outpatient, office and remote visits ($273). Compared to patients with lowest DMC savings (bottom-quartile), those with highest DMC savings (top-quartile) were detected 3.7-months (mean) earlier, corresponding to approximately $13,540 PPPM greater DMC saved.
CONCLUSIONS: Earlier detection of PAH facilitated by this ML algorithm was associated with meaningful savings in DMC, indicating that such tools, when implemented within health systems, may support downstream resource allocation and cost savings.
Conference/Value in Health Info
2026-05, ISPOR 2026, Philadelphia, PA, USA
Value in Health, Volume 29, Issue S6
Code
EE68
Topic
Economic Evaluation
Topic Subcategory
Cost/Cost of Illness/Resource Use Studies
Disease
No Additional Disease & Conditions/Specialized Treatment Areas, SDC: Rare & Orphan Diseases