Using machine learning to improve the readability of hospital discharge instructions for heart failure

BackgroundLow health literacy is associated with poor health outcomes. Hospital discharge instructions are often written at advanced reading levels, limiting patients with low health literacy ability to follow medication instructions or complete other necessary care. Previous research demonstrates that improving the readability of discharge instructions reduces hospital readmissions and decreases healthcare costs. We aimed to use artificial intelligence (AI) to improve the readability of discharge instructions. Methodology/Principal FindingsWe collected a series of discharge instructions for adults hospitalized for heart failure (n=423), which were then manually simplified to a lower reading level to create two parallel sets of discharge instructions. Only 343 sets were then processed via AI-based machine learning to create a trained algorithm. We then tested the algorithm on the remaining 80 discharge instructions. Output was evaluated quantitatively using Simple Measure of Gobbledygook (SMOG) and Flesch-Kincaid readability scores and cross-entropy analysis and qualitatively. Using this test dataset (n=80), the average reading levels were: original discharge instructions (SMOG: 10.5669{+/-}1.2634, Flesch-Kincaid: 8.6038{+/-}1.5509), human-simplified instructions (SMOG: 9.4406{+/-}1.0791, Flesch-Kincaid: 7.2221{+/-}1.3794), and AI-simplified instructions (SMOG: 9.3045{+/-}0.9531, Flesch-Kincaid: 7.0464{+/-}1.1308). AI-simplified instructions were significantly different from original instructions (p<0.00001). The algorithm made appropriate changes in 26.1% of instances to the original discharge instructions and improved average reading levels by 1.26{+/-}0.32 grade levels (SMOG) and 1.02{+/-}0.47 grade levels (Flesch-Kincaid). Cross-entropy analysis showed that as the data set increased in size, the function of the algorithm improved. Conclusions/SignificanceThe AI-based algorithm learned meaningful phrase-level simplifications from the human-simplified discharge instructions. The AI simplifications, while not in complete agreement with the human simplifications, do appear as statistically significant improvements to SMOG and Flesch-Kincaid reading levels. The algorithm will likely produce more meaningful and concise simplifications among discharge instructions as it is trained on more data. This study demonstrates an important opportunity for AI integration into healthcare delivery to address health disparities related to limited health literacy and potentially improve patient health. Author summaryPatient-facing materials are often written at too high of a reading level for patients, such as hospital discharge instructions. These instructions provide critical information on how to control health conditions, take medications, and attend follow-up visits. Difficulty understanding these instructions could lead to the patient returning to the hospital if they do not understand how to control their health condition. Improving the readability of discharge instructions can reduce hospital readmissions. It may improve health outcomes for patients and reduce healthcare costs. Artificial intelligence (AI) may be used to improve the reading level of patient-facing materials. Our work aims to create a tool that can accomplish this goal. We obtained hospital discharge instructions for heart failure. Discharge instructions were edited by medical experts to improve their readability. This created two sets of discharge instructions that were processed using AI. We created and tested an AI tool to automatically simplify discharge instructions. Although not perfect, we found that the tool was successful. This research shows that AI can be used to address health literacy needs within health care by making patient-facing health materials easier to understand. This is important to empower all patients to take action to improve their health.