| The Electronic Health Record(EHR)is an electronic version of patient medical records.Compared with traditional manual records,EHR systems can automatically collect,store and manage patients’ medical information,greatly reducing the workload of healthcare professionals.With the development of artificial intelligence technology,the era of medical informatization has arrived.Artificial intelligence algorithms have been widely used in the medical field,such as the use of EHR for diagnosis and treatment,etc.However,there are still problems such as poor interpretability,incomplete diagnosis,and optimization of medication strategies.Especially in ICU,the patient’s death risk is a key issue that ICU clinicians need to pay attention to all the time,and designing efficient and interpretable death risk prediction algorithms can improve the patient’s prognosis.ICU clinicians can only treat patients after they have an overall understanding of the patient’s death risk.When ICU patients suffer from multiple diseases,the primary and secondary order of the diseases affects the clinician’s treatment strategy and the patient’s medical reimbursement.The diagnostic algorithm of the primary and secondary order can help ICU clinicians to make a comprehensive diagnosis and reduce the leakage and misdiagnosis of the diseases.After making a comprehensive diagnosis,ICU clinicians focus on treating the diseases that have the greatest impact on the patient,which can maximize the improvement of the patient’s physiological state.But there is usually a problem of under-medication or overmedication in this process.Based on the problems of patient death risk,disease diagnosis,and treatment strategies encountered in the ICU patient diagnosis and treatment process.This dissertation takes ICU patients and diabetic ketoacidosis(DKA)patients in ICU as examples,and focus on the EHR-based algorithms for death risk and disease diagnosis and treatment in three aspects: death risk prediction of ICU patients,ICU patients’ disease primary and secondary order diagnosis,and ICU patients’ treatment decision optimization.The specific work is as follows:First,this dissertation proposes an EHR-based algorithm for predicting the interpretable mortality risk of ICU patients to solve the problems of category imbalance in EHR and the non-interpretability of generative models.To address these two problems,Conditional Wasserstein Generative Adversarial Network(CWGAN)and Shapley value interpretable methods are investigated,respectively.The category imbalance problem in EHR is solved by generating small categories of ICU death patient samples using CWGAN.The impact of the additional data generated by CWGAN on the classification model is analyzed using the Shapley value method.Compared with the traditional SMOTE generation model,CWGAN generates data of higher quality,and its generated samples are more consistent with the original data distribution.In the mortality risk prediction task,CWGAN-generated data significantly outperformed the traditional generative model for the classification task metrics.For the inexplicability of the generative model,an interpretable method based on the Shapley value is used to analyze the key physiological variables that lead to the deaths of ICU patients from both ICU patients in the validation set and DKA patients in the ICU,respectively.The interpretability of the generative model is analyzed with the additional generative data from CWGAN.Finally,the limitations of AI model interpretability are discussed.Secondly,this dissertation proposes an EHR-based disease primary and secondary order diagnosis algorithm to solve the problem of disease primary and secondary order of ICU patients.In the process of disease diagnosis,the algorithm based on natural language processing and classifier chain can assist ICU clinicians in diagnosis and help them to identify the primary and secondary order of diseases quickly and efficiently.It can assist them in manual coding of disease cases in the DRG(Diagnosis Related Groups)system and so on.In the case of ICU patients suffering from multiple diseases,the primary and secondary order of diseases can provide appropriate diagnosis for patients with different disease sequences and reduce misdiagnosis and omission of diseases.By utilizing the ICU clinical medical text in EHR and adopting text embedding technology,this dissertation designs a diagnosis algorithm based on classifier chain to analyze the sequentiality of different diseases on patients,so as to realize the diagnosis of the primary and secondary order of diseases.This dissertation takes the ICU patients and the DKA patients as examples to analyze the secondary diseases that are more likely to be suffered from ICU,so as to make the diagnosis of the ICU clinicians more comprehensive and efficient.Finally,this dissertation proposes an EHR-based treatment decision optimization algorithm for ICU patients to optimize the treatment decision(insulin)of DKA patients in ICU.In clinical treatment decision-making,the physiological state of the patient will change accordingly with the treatment strategy given by the clinicians.There are cases of underdosing or overdosing by the clinician,and the decision-making of medication is directly related to the prognosis of the patient.The main treatment for patients with DKA is the injection of clinical insulin.Reinforcement learning algorithm can be used to find the optimal insulin dosing strategy,thus providing decision-making assistance to ICU clinicians.In addition,considering that there are scenarios in which multiple clinicians discuss and make decisions on treatment strategies in real medical environments,which can reduce individual bias and improve decision-making level.Aiming at the scenario of multi-clinicians treatment,a multi-agent value decomposition reinforcement learning algorithm is designed to fuse the decision-making of multi-agent through weighting.This algorithm can effectively simulate the decision-making of multiple clinicians,and successfully optimize the decision-making of clinicians.For the problem of death risk and disease diagnosis and treatment of ICU patients,this dissertation designs corresponding algorithms based on EHR to form an integrated dissertation of death risk and disease diagnosis and treatment mainly for ICU patients.The dissertation takes DKA patients in ICU as a special case to illustrate the help of the algorithm on specific diseases and show the universality of the algorithm.The algorithm proposed in this dissertation effectively utilizes artificial intelligence technology to solve the problems of patient death risk,disease primary and secondary order,and medication strategy encountered in the diagnosis and treatment of ICU patients.It helps to improve the decision-making level of ICU clinicians and reduce the workload of healthcare workers.The results in this dissertation have wide applicability and can also be applied to other patients outside the ICU,which will hopefully have a positive impact and help more patients. |
