OBJECTIVE: This study explores the feasibility of using machine learning to predict accurate versus inaccurate diagnoses made by pathologists based on their spatiotemporal viewing behavior when evaluating digital breast biopsy images. MATERIALS AND METHODS: The study gathered data from 140 pathologists of varying experience levels who each reviewed a set of 14 digital whole slide images of breast biopsy tissue. Pathologists' viewing behavior, including zooming and panning actions, was recorded during image evaluation. A total of 30 features were extracted from the viewing behavior data, and 4 machine learning algorithms were used to build classifiers for predicting diagnostic accuracy. RESULTS: The Random Forest classifier demonstrated the best overall performance, achieving a test accuracy of 0.81 and area under the receiver-operator characteristic curve of 0.86. Features related to attention distribution and focus on critical regions of interest were found to be important predictors of diagnostic accuracy. Further including case-level and pathologist-level information incrementally improved classifier performance. DISCUSSION: Results suggest that pathologists' viewing behavior during digital image evaluation can be leveraged to predict diagnostic accuracy, affording automated feedback and decision support systems based on viewing behavior to aid in training and, ultimately, clinical practice. They also carry implications for basic research examining the interplay between perception, thought, and action in diagnostic decision-making. CONCLUSION: The classifiers developed herein have potential applications in training and clinical settings to provide timely feedback and support to pathologists during diagnostic decision-making. Further research could explore the generalizability of these findings to other medical domains and varied levels of expertise.

MOTIVATION: Permutation testing is very popular for analyzing microarray data to identify differentially expressed (DE) genes; estimating false discovery rates (FDRs) is a very popular way to address the inherent multiple testing problem. However, combining these approaches may be problematic when sample sizes are unequal. RESULTS: With unbalanced data, permutation tests may not be suitable because they do not test the hypothesis of interest. In addition, permutation tests can be biased. Using biased P-values to estimate the FDR can produce unacceptable bias in those estimates. Results also show that the approach of pooling permutation null distributions across genes can produce invalid P-values, since even non-DE genes can have different permutation null distributions. We encourage researchers to use statistics that have been shown to reliably discriminate DE genes, but caution that associated P-values may be either invalid, or a less-effective metric for discriminating DE genes.

BACKGROUND: As part of its broad and ambitious mission, the MicroArray Quality Control (MAQC) project reported the results of experiments using External RNA Controls (ERCs) on five microarray platforms. For most platforms, several different methods of data processing were considered. However, there was no similar consideration of different methods for processing the data from the Agilent two-color platform. While this omission is understandable given the scale of the project, it can create the false impression that there is consensus about the best way to process Agilent two-color data. It is also important to consider whether ERCs are representative of all the probes on a microarray. RESULTS: A comparison of different methods of processing Agilent two-color data shows substantial differences among methods for low-intensity genes. The sensitivity and specificity for detecting differentially expressed genes varies substantially for different methods. Analysis also reveals that the ERCs in the MAQC data only span the upper half of the intensity range, and therefore cannot be representative of all genes on the microarray. CONCLUSION: Although ERCs demonstrate good agreement between observed and expected log-ratios on the Agilent two-color platform, such an analysis is incomplete. Simple loess normalization outperformed data processing with Agilent's Feature Extraction software for accurate identification of differentially expressed genes. Results from studies using ERCs should not be over-generalized when ERCs are not representative of all probes on a microarray.

BACKGROUND: There are currently many different methods for processing and summarizing probe-level data from Affymetrix oligonucleotide arrays. It is of great interest to validate these methods and identify those that are most effective. There is no single best way to do this validation, and a variety of approaches is needed. Moreover, gene expression data are collected to answer a variety of scientific questions, and the same method may not be best for all questions. Only a handful of validation studies have been done so far, most of which rely on spike-in datasets and focus on the question of detecting differential expression. Here we seek methods that excel at estimating relative expression. We evaluate methods by identifying those that give the strongest linear association between expression measurements by array and the "gold-standard" assay. Quantitative reverse-transcription polymerase chain reaction (qRT-PCR) is generally considered the "gold-standard" assay for measuring gene expression by biologists and is often used to confirm findings from microarray data. Here we use qRT-PCR measurements to validate methods for the components of processing oligo array data: background adjustment, normalization, mismatch adjustment, and probeset summary. An advantage of our approach over spike-in studies is that methods are validated on a real dataset that was collected to address a scientific question. RESULTS: We initially identify three of six popular methods that consistently produced the best agreement between oligo array and RT-PCR data for medium- and high-intensity genes. The three methods are generally known as MAS5, gcRMA, and the dChip mismatch mode. For medium- and high-intensity genes, we identified use of data from mismatch probes (as in MAS5 and dChip mismatch) and a sequence-based method of background adjustment (as in gcRMA) as the most important factors in methods' performances. However, we found poor reliability for methods using mismatch probes for low-intensity genes, which is in agreement with previous studies. CONCLUSION: We advocate use of sequence-based background adjustment in lieu of mismatch adjustment to achieve the best results across the intensity spectrum. No method of normalization or probeset summary showed any consistent advantages.