Step-by-Step Guidance to Differential Anemia Diagnosis with Real-World Data and Deep Reinforcement Learning

Lillian Muyama, Estelle Lu, Geoffrey Cheminet, Jacques Pouchot, Bastien Rance, Anne-Isablle Tropeano, Antoine Neuraz, Adrien Coulet · 2024

Read Step-by-Step Guidance to Differential Anemia Diagnosis with Real-World Data and Deep Reinforcement Learning by Lillian Muyama, Estelle Lu, Geoffrey Cheminet, Jacques Pouchot, Bastien Rance, Anne-Isablle Tropeano, Antoine Neuraz, Adrien Coulet (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-02273.txt.

Introduction

Lillian Muyama and co-authors build step-by-step guidance for differential diagnosis of anemia using real-world data and deep reinforcement learning. Their 2024 paper takes inspiration from clinical guidelines, which spell out the questions to answer.

Their model learns from electronic health records to pick the best sequence of actions toward a diagnosis. Algorithms were tested on a synthetic dataset built from expert-defined pathways, and on real-world data. Performance was checked across several scenarios. The learning methods matched strong baselines while adding a key benefit: they generate the diagnostic pathway step by step. That transparency lets a clinician follow the reasoning, not just the final label.

Catalog record

CREATORLillian Muyama, Estelle Lu, Geoffrey Cheminet, Jacques Pouchot, Bastien Rance, Anne-Isablle Tropeano, Antoine Neuraz, Adrien Coulet
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE50 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-02273

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Step-by-Step Guidance to Differential Anemia Diagnosis with Real-World Data and Deep Reinforcement Learning Lillian Muyama, Estelle Lu, Geoffrey Cheminet, Jacques Pouchot, Bastien Rance, Anne-Isablle Tropeano, Antoine Neuraz, Adrien Coulet Abstract Clinical diagnostic guidelines outline the key questions to answer to reach a diagnosis. Inspired by guidelines, we aim to develop a model that learns from electronic health records to determine the optimal sequence of actions for accurate diagnosis. Focusing on anemia and its sub-types, we employ deep reinforcement learning (DRL) algorithms and evaluate their performance on both a synthetic dataset, which is based on expert-defined diagnostic pathways, and a real-world dataset. We investigate the performance of these algorithms across various scenarios. Our experimental results demonstrate that DRL algorithms perform competitively with state-of-the-art methods while offering the significant advantage of progressively generating pathways to the suggested diagnosis, providing a transparent decision-making process that can guide and explain diagnostic reasoning. 1 Introduction Clinicians often rely on Clinical Practice Guidelines (CPGs) to make informed diagnostic decisions. CPGs are systematically developed recommendations to assist healthcare providers in choosing suitable healthcare strategies tailored to specific clinical scenarios [ 1 ] . The primary purpose of these guidelines is to standardize and rationalize clinical decisions. For diagnostic purposes, the recommended actions may include physical examinations of a patient, laboratory tests, imaging studies, and other relevant assessments. However, despite their undeniable role in clinical decision-making, CPGs have several limitations. First, their development requires input from a diverse panel of experts, leading to a time-consuming and costly process. This implies that updates to CPG take several years, which makes them unsuitable for rapidly evolving medical practices, such as the introduction of new diagnostic tests or the emergence of a new disease. Second, due to the substantial resources needed to develop these guidelines, it is not practical to create CPGs for every known medical condition. Third, guidelines are generally designed to address the needs of the majority of the population, potentially overlooking rare conditions or uncommon populations. Additionally, the adoption of CPGs in real-world clinical settings is often complicated by various factors, such as outdated recommendations, clinicians’ reliance on personal knowledge and experience over guidelines, and resource constraints. In response to these challenges, several studies have explored innovative approaches for learning clinical pathways from data. These methods have the potential to complement guidelines by providing insights where guidelines are absent, incomplete, or unsuitable. The vast amount of data collected during care, particularly through Electronic Health Records (EHRs), presents an opportunity to enhance clinical decision-making. EHRs include diverse patient information, including laboratory test results, medications, examinations, and diagnoses, reflecting clinical practice. Following directions proposed by [ 2 ] , our study develops a data-driven methodology that supports the clinical decision-making process in a step-by-step manner. We believe this approach could reduce unnecessary tests, optimize costs, and offer personalized and accurate diagnoses, particularly for uncommon and overlooked patients. In this work, we build on a previous study [ 3 ] , where authors used a Deep Reinforcement Learning (DRL) approach to learn diagnostic decision pathways for anemia and systemic lupus erythematosus. The authors created synthetic datasets, which were labeled based on existing CPGs, and manually added noise and missing data to test the robustness of their approach. In the present study, we go further, but focus only on anemia, for three main reasons: its diagnosis is primarily based on a series of laboratory tests available in most EHRs; it is a common diagnosis, implying that the associated amount of data is sufficient to train ML models; and the differential diagnosis of anemia is often complex, making its guidance particularly useful. Our approach proceeds as follows: First, we collaborated with a clinician specialized in internal medicine to develop a diagnostic decision tree (DT) for anemia, which we name the “expert-defined” DT. This tree served as the basis for generating a synthetic dataset, on which we applied our method to generate baseline diagnostic pathways. Second, we tested our approach on a real-world dataset, under three distinct scenarios: ( i ) applying the model trained solely on synthetic data to the real-world data; ( ii ) fine-tuning the synthetic data-trained model with a subset of the real-world data; and ( iii ) training a new model from scratch using a subset of the real-world dataset. Our main contributions are: 1) We demonstrate that deep reinforcement learning methods have the potential to provide step-by-step guidance for clinical diagnosis in a real-world setting. 2) We show that in cases of insufficient data, the use of a synthetic dataset may improve model performance. 2 Related Work The discovery of clinical pathways from patient data has been widely explored using mostly unsupervised learning techniques such as process mining and topic modeling. For instance, [ 4 ] employed process mining to discover clinical pathways for sepsis, [ 5 ] applied K-medoids and logistic regression to identify pathways to reduce hospitalizations, and [ 6 ] combined process mining with topic modeling to generate clinical pathways for intracranial hemorrhage. Other frequently employed unsupervised learning methods include hidden Markov models [ 7 , 8 ] and sequence mining [ 9 , 10 ] . However, these approaches often require that the trajectory (sequence of clinical events) of each patient is known, which can be problematic in real-world settings where event timings may overlap, be delayed, or recorded unreliably. Additionally, several studies have leveraged machine learning methods to aid clinical decision-making by predicting patient outcomes [ 11 , 12 ] , diagnoses [ 13 , 14 ] , among other tasks. However, unlike these works, our study aims not only to provide diagnostic predictions but also to detail the steps taken to reach that diagnosis, thereby mimicking the decision-making process of a clinician. Previous studies have applied reinforcement learning (RL) methods for costly feature acquisition in classification tasks [ 15 , 16 ] . However, these works did not specifically focus on clinical diagnosis. RL Methods have also been utilized to develop optimal dosing strategies in various domains, including chemotherapy [ 17 , 18 ] , radiotherapy [ 19 , 20 ] , and glucose control [ 21 , 22 ] . These applications can be viewed as developing treatment pathways for patients. Several works [ 23 , 24 , 25 ] built dialogue systems, where patients were queried about their symptoms, eventually leading to a diagnosis. While our approach similarly formulates the diagnosis process as a sequential decision-making problem using DRL, these studies rely on self-reported symptoms. In contrast, we aim to use EHRs, which we believe offer a robust alternative as they encompass data routinely collected in clinical practice, including objective and standardized measurements such as laboratory results. In another related work [ 26 ] , the authors sought to optimize the financial cost of several clinical processes, including the prediction of Acute Kidney Injury using DRL. While they were successful in reducing laboratory test costs compared to other models, the sequences of actions leading to these reductions and subsequent diagnoses were not discussed or shown. This omission neglects the relevance and explainable dimension of pathways. 3 Methods 3.1 Decision Problem We consider the clinical diagnosis process as a sequential decision-making problem and formulate it as a Markov Decision Process (MDP)

Read the complete text