عنوان مقاله:
هوش مصنوعی در دانش قلب شناسی (کاردیولوژی)
Artificial Intelligence in Cardiology
سال انتشار: 2018
رشته: مهندسی کامپیوتر - پزشکی
گرایش: هوش مصنوعی - انفورماتیک پزشکی - قلب و عروق
دانلود رایگان این مقاله:
دانلود مقاله هوش مصنوعی در قلب شناسی
مشاهده سایر مقالات جدید:
A brief survey of supervised machine learning algorithms in cardiology
Ultimately, supervised machine learning is the attempt to model how independent variables relate to a dependent variable (Table 1). In machine learning, one must choose a strategy (by selecting a particular algorithm) to discover these relationships. This section highlights several algorithms that may be used in cardiovascular settings and provides a summary of supervised and unsupervised algorithms. REGULARIZED REGRESSION. Imagine a situation where you have dozens, hundreds, or thousands of variables collected on just a few patients. You wish to decipher which variables are actually of relevance. As noted earlier, in this “big p, small n” (i.e., large number of features relative to a small number of samples) situation, a potential solution is a class of techniques called regularized regression. In this context, regularization means the introduction of additional constraints to decrease model complexity, thus allowing the model to generalize better to other data. Some common forms of regularized regression are called LASSO (least absolute shrinkage and selection operator) regression, ridge regression, and elastic net regression. To give an example of the benefits of regularization, Halim et al. (20) used logistic elastic net regression when combining proteomic measurements with clinical variables to predict the incidence of myocardial infarction or death. As these investigators noted, elastic net regression allowed them to evaluate many independent features in a novel way and ultimately helped them find candidate proteomic biomarkers for cardiovascular event risk.
Reinforcement learning
Reinforcement learning algorithms learn behavior through trial and error given only input data and an outcome to optimize (Figure 3D). In popular culture, a breakthrough example from 2015 highlights the power of this technique. A group of researchers trained a reinforcement learning model on a variety of classic Atari 2600 video games and provided only video input and the game’s final score (47). The model “learned” the optimal method to maximize the final score. More recently, a research group at Google trained a reinforcement learning model to beat a world champion at the Chinese board game Go, a task once believed too difficult for computers (48). Specifically, reinforcement learning algorithms consist of an agent at a particular time interacting with an environment. An action is selected for each time point according to some selection policy. Transitions to the next state are then performed, and a reward is received depending on the result of the transition. The restricted learning model aims to maximize the expectation of long-term rewards from each state visited. Application of reinforcement learning to health care and cardiology thus far has been scarce. Socalled dynamic treatment regimens that tailor treatment decisions to a patient’s characteristics are potential applications for reinforcement learning algorithms because of their inherent sequential decision-making structure, although statistical causal inference approaches also show promise when applied to this problem (49). Work from Shortreed et al. (50) demonstrated that reinforcement learning can work for optimization of treatment policies in chronic illnesses. Importantly, these investigators showed that reinforcement learning can overcome the problem of missing data and quantify the uncertainty of recommended policy. More recent work using restricted learning to manage weaning of mechanical ventilation in intensive care units shows great promise in minimizing rates of reintubation and regulating physiological stability (51). We envision that reinforcement learning models will eventually be commonplace and function as physician extenders in day-to-day clinical practice, either built into the EHRs or as part of devices worn by the clinician.
چکیده
هوش مصنوعی و یادگیری ماشینی، می توانند تقریبا بر روی تمام ابعاد شرایط انسان تاثیر داشته باشند و دانش قلب شناسی هم از این قاعده مستثنی نیست. این مقاله یک راهنمای خوب برای متخصص های بالینی در رابطه با ابعاد مربوط به هوش مصنوعی و یادگیری ماشینی فراهم می کند و بعضی از کاربرد های انتخابی این روش را در قلب شناسی تا به امروز شناسایی می کند و این موضوع را بررسی می کند که داروهای قلبی عروقی چطور میتوانند از هوش مصنوعی در آینده استفاده کنند. به صورت خاص، این مقاله نخست مفاهیم مدل سازی پیش بینی را که مرتبط با قلب شناسی می باشد را شناسایی می کند مانند گزینش ویژگی و مشکلات رایج مانند دوشعبه سازی نا مناسب. دوم این که ما در این مقاله الگوریتم های رایج مورد استفاده در یادگیری با سرپرست را بررسی می کنیم و کاربرد های انتخاب شده را در قلب شناسی و زمینه های علمی مربوطه، بررسی می کنیم. سوم، ما پیشرفت یادگیری عمقی و روش های مربوطه که به صورت کلی با نام یادگیری بدون سرپرست شناخته می شود را بررسیمی کنیم که نمونه های زمینه ای را در پزشکی عمومی و پزشکی قلب و عروق ارائه می کند و سپس نشان می دهیم که این روش ها را چطور می توان مورد استفاده قرار داد تا موجب بهبود دانش قلب و عروق شد و خروجی های به دست آمده را ارتقا داد.
