Statistical agencies rely on sampling techniques to collect socio-demographic data crucial for policy-making and resource allocation. This paper shows that surveys of important societal relevance introduce sampling errors that unevenly impact group-level estimates, thereby compromising fairness in downstream decisions. To address these issues, this paper introduces an optimization approach modeled on real-world survey design processes, ensuring sampling costs are optimized while maintaining error margins within prescribed tolerances. Additionally, privacy-preserving methods used to determine sampling rates can further impact these fairness issues. This paper explores the impact of differential privacy on the statistics informing the sampling process, revealing a surprising effect: not only is the expected negative effect from the addition of noise for differential privacy negligible, but also this privacy noise can in fact reduce unfairness as it positively biases smaller counts. These findings are validated over an extensive analysis using datasets commonly applied in census statistics.

Neural networks (NNs) can learn to rely on spurious signals in the training data, leading to poor generalisation. Recent methods tackle this problem by training NNs with additional ground-truth annotations of such signals. These methods may, however, let spurious signals re-emerge in deep convolutional NNs (CNNs). We propose Targeted Activation Penalty (TAP), a new method tackling the same problem by penalising activations to control the re-emergence of spurious signals in deep CNNs, while also lowering training times and memory usage. In addition, ground-truth annotations can be expensive to obtain. We show that TAP still works well with annotations generated by pre-trained models as effective substitutes of ground-truth annotations. We demonstrate the power of TAP against two state-of-the-art baselines on the MNIST benchmark and on two clinical image datasets, using four different CNN architectures.

Background: Brain tumors are the leading cause of cancer death in the under-40s. The commonest malignant brain tumor is a Glioblastoma multiforme (GBM) with less than 5% 5-year survival. Low skeletal muscle mass is associated with poor survival in cancer and is measurable on routine imaging, but manual muscle mass quantification is time-consuming and susceptible to interrater inconsistency. In patients with brain tumors, measurement of the thickness of the temporalis muscle acts as a proxy. We present a fully-automated deep learning-based system for temporalis muscle quantification, a skeletal muscle mass surrogate on MRI head. Methods: MRI scans of 330 patients were obtained from four different datasets. Two 2D U-Nets were trained, one to segment the eyeballs and the other to segment the temporalis muscle, and used to quantify the cross-sectional area of the eyeballs and temporalis muscle. The eyeball segmentation was used to chose a consistent level on which to assess temporalis muscle mass. We assessed accuracy of segmentation using DICE and Hausdorff scores, and assessed the performance of the system to choose the correct slice on the MRI by comparing it with a manual choice of MRI slice. Results: The set models predict eyeball and temporalis muscle segmentations with good accuracy. Mean Dice scores is 0.90±0.03 and 0.90±0.05 and Hausdorff distances are 2.88±0.60 and 1.89±0.35, respectively. The automatic pipeline chooses slices for segmentation that are either identical or close to the manual choice in 96.1% of cases. Conclusions: We have developed an end-to-end system that uses two independently trained U-Nets to first segment the eyeball and then use that as a reference point to pick the correct slice of the MRI on which to use the second U-Net to measure temporalis cross-sectional area. This allows for the automated processing of head MRI scans to measure temporalis muscle mass, which has been previously shown to correlate with body muscle mass and survival in several cancer types.

Survival prediction is a key task in medicine. Existing models are based on statistical techniques, such as the Cox models and there is limited work on the application of machine learning. In this paper we demonstrate that the K-Nearest Neighbour algorithm can be used for survival prediction. We show that its performance is as good as that of standard techniques, and that it provides a clear interpretation of the results. We show that pre-processing methods improve performance, and evaluate the performance across 20 different datasets with differing properties to show that the model performs well under various conditions. For low event rate datasets we show that KNN can outperform the Cox model.