Study On The Prediction Model Of Levothyroxine Medication In Patients With Hypothyroidism

Objective: The incidence of hypothyroidism in china remains high.During the process of diagnosis and treatment of hypothyroidism in clinical practice,the therapeutic dose of levothyroxine（LT4）is affected by various factors,and patients often experience insufficient or excessive substitution,Which is not conducive to treatment progress.Therefore,in order to avoid excessive or insufficient replacement of LT4 in patients and restore thyroid function as soon as possible,the LT4 treatment regimen should fully realize individualized dosing.In this study,machine learning algorithm was used to establish LT4 individualized drug delivery model based on real-world data,in order to optimize LT4 treatment regimen for patients with hypothyroidism and improve the compliance rate of hypothyroidism treatment.Methods: Firstly,searching and analyzing the published research literature systematically on the influencing factors of LT4 medication in patients with hypothyroidism,we fully gained the possible influencing factors of establishing a prediction model,and made a real-world data extraction catalog.Secondly,we collected the medical data of inpatients with hypothyroidism in Sichuan Provincial People’s Hospital from January 1,2016 to September 30,2021,screened and extracted the target case data through the established inclusion and exclusion criteria,and then obtain the research data set through preliminary preprocessing.We used traditional mathematical statistics methods to statistically describe the variables in the data set,and discussed the distribution characteristics of variables in groups.Finally,five machine learning algorithms,that is random forest,support vector machine,gradient boosting decision tree,extreme gradient boosting and Adaboost,were used to establish a prediction model for whether thyroid stimulating hormone（TSH）returned to normal after treatment and a prediction model for LT4 dose.Ultimately,we selected the model with the best performance through the corresponding model evaluation index.Results: The influencing factors from included in the literature study can be summarized into four parts: patient’s basic information,comorbidities information,medication information and related examinations.After the real-world data was screened by inclusion and exclusion criteria,a research data set with a sample size of1154 was finally obtained.According to the TSH level,690 were within the normal range,62 were low,and 402 were high.After preliminary preprocessing,a total of 138 variables were obtained,including 49 qualitative variables and 89 quantitative variables.Before feature screening,the best-performing TSH classification model was obtained by the GBDT algorithm,and its AUC was 0.762;the best-performing LT4 dose prediction model was obtained by the Ada Boost algorithm,and its MAE and MSE were 0.338 and0.204,respectively.After feature screening,the TSH classification prediction model obtained by the SVM algorithm had the best performance,and its AUC was 0.770;the LT4 dose prediction model obtained by the SVM algorithm had the best performance,and its MAE and MSE were 0.331 and 0.196,respectively.Conclusion: The machine learning algorithm can be used to predict whether TSH returns to normal and the dose of LT4 in patients with hypothyroidism.The established predictive model has a certain reference value,which provides a new idea for the formulation of treatment plans for patients with hypothyroidism.Also,data mining research on chronic diseases provides methodological reference on other chronic disease.