Content: When you are presented with topic, such as machine learning, that is not familiar, you certainly can appreciate how difficult is must be to appreciate clinical lectures, when this is all new. In the current case, it is with great appreciation that I review the blogs of my classmates, who show a significant understanding of this topic, essentially the methods of handling and analyzing data or data mining with various approaches. The article suggested by Annie is instructive in this regard and states that “classification is a data mining task that can use various statistical or machine learning algorithms to assign instances to different classes.” In fact, one of the foundations of clinical decision support is the sequence of machine learning by various approaches. Whether or not the Intern program for clinical evaluation and diagnosis is a form of machine learning or smart computer analysis is not completely clear to me. Greenes in his book documents a number of limitations and states: “most clinical decision support systems in current use do not learn from data and still rely on rule based paradigms.” There has been limited adoption by the biomedical community of techniques for determining data patterns and relationships (Greenes, 2006), which is the purpose of data mining and machine learning. Clinicians do not necessarily understand probabilistic models for generating and analyzing knowledge and we have seen the limitations of the Bayesian Theorems in clinical decision making. Would it be good or not so good to plug data in from different sources and different formats and get out analysis. However, this is highly dependent on the designs of machines and algorithms and programs by us. “Algorithms that can recognize regularities in data and can construct a model” can be utilized. “Data mining techniques are pattern recognition techniques intended to find correlations and relationships in plethora of data.” Greenes goes on to distinguish supervised versus unsupervised systems. The latter represent exploratory analysis without predefined hypotheses or conclusions, unveiling potential “hidden patterns.” These are not applicable to clinical situations at this time, but supervised models can be applied to clinical decision making to develop predictive models from data handling. In this regard, machine learning techniques such as classification trees and artificial neural networks have been most useful, particularly in the latter in critical care situations, i.e. in the ICU. However, from a clinical standpoint most of these models are research models rather than clinical models (Greenes). It is therefore instructive to start to evaluate some of the other models/methods of machine learning, discussed by Mr. Li and start to think about their utility in bioinformatics and also in biomedical informatics. “Machine learning refers to changes in systems that are concerned with artificial intelligence…recognition, robot control, diagnosis, planning, prediction.” (Nilsson, 1996). One wonders if the equipment and paradigms that Kanav is working on in the operating room would be classified as machine learning based on human interaction. Inputs are viewed as attributes and outputs as classes. I do not have enough knowledge on these subjects to determine how raw or organized medical data in different formats can be applied to algorithms that might be created by K-nearest neighbor classifier or support vector machines but I can see clinical application for decision tree analysis and naïve bayes classifier, as well as neural networks mentioned above. A key question is how the hypotheco-deductive approaches and other clinical decision paradigms might be linked to and based on the machine learning and data mining approaches above. It would be useful to hear other’s ideas on this. I suspect that K nearest neighbor and support vector protocols would be focused on unsupervised rather than supervised situations. But Support vector machines do fit data by lines and might apply to supervised situations. Neural networks, as defined by Nilsson, are dependent on non-linear threshold functions, similar to the way that nerve turn on may required threshold elicitation. I am not sure that this simplistic concept means that mathematical model can simulate the nervous system. The second lecture by Dr. Pettiti was succinct and quite useful for me, as a researcher. Our primary focus has been on experimental studies, but the plan is to move into epidemiological studies that are directly related to our animal models of human disease.

Posted by  Stuart