《学术论文集锦[开错]》紫微客

16、016

　　4.3.5 Other defects and improvements

　　In the innovative design of Part 4.2, there is an another flaw, that is the criteria for patient classification (the number of qualified indicators ≥ 2) is decided based on previous experience, but perhaps this is not the most reliable scenario. For example, when we judge the sensitivity of cells to drugs, based on the principle of using the lowest dose to achieve the optimal results, perhaps the most noteworthy point should be the cell inhibition rate at low concentrations. For this consideration, I think it is necessary to introduce the “weight theory”, that is, the groups with lower concentration have higher weight, such as A (50%), B (30%), C (20%), D (10%), and then classify patients (Low sensitivity, Medium sensitivity, High sensitivity).

　　In addition, we can also improve the design of the cluster analysis. In Part 4.2, I take cell inhibition rate as variable characteristic required for cluster analysis, but it might be better if conducting cluster analysis according to the corresponding fitting curve of each experimental group. The specific method is to take the slope of two adjacent points in each line, and then take these slopes as the variable feature to perform the cluster analysis. At the same time, we can also consider comparing the degree of linear correlation between groups using Pearson correlation coefficient by calculating the vectors formed by 8 points (concentration vs inhibition rate).

　　After that, the evaluation indicator of clustering effectiveness can use Davies-Bouldin index (DB), which is an index proposed by Davies and Bouldin to evaluate the quality of the clustering algorithm[28]. The smaller the value of DB index indicates that the better the clustering effect, and its mathematical expression is as follows: