21 June 2024 in IEEE Journal of Biomedical and Health Informatics
Variational Inference (VI) is a commonly used technique for approximate Bayesian inference and uncertainty estimation in deep learning models, yet it comes at a computational cost, as it doubles the number of trainable parameters to represent uncertainty. This rapidly becomes challenging in high-dimensional settings and motivates the use of alternative techniques for inference, such as Monte Carlo Dropout (MCD) or Spectral-normalized Neural Gaussian Process (SNGP). However, such methods have seen little adoption in survival analysis, and VI remains the prevalent approach for training probabilistic neural networks. In this paper, we investigate how to train deep probabilistic survival models in large datasets without introducing additional overhead in model complexity. To achieve this, we adopt three probabilistic approaches, namely VI, MCD, and SNGP, and evaluate them in terms of their prediction performance, calibration performance, and model complexity. In the context of probabilistic survival analysis, we investigate whether non-VI techniques can offer comparable or possibly improved prediction performance and uncertainty calibration compared to VI. In the MIMIC-IV dataset, we find that MCD aligns with VI in terms of the concordance index (0.748 vs. 0.743) and mean absolute error (254.9 vs. 254.7) using hinge loss, while providing C-calibrated uncertainty estimates. Moreover, our SNGP implementation provides D-calibrated survival functions in all datasets compared to VI (4/4 vs. 2/4, respectively). Our work encourages the use of techniques alternative to VI for survival analysis in high-dimensional datasets, where computational efficiency and overhead are of concern.
3 February 2022 in SN Computer Science
30 July 2024 in Journal of Public Health
Aim: The present study has designed, implemented, and evaluated a machine learning model that can predict fall risk and fall occurrence in community-dwelling elderly based on their home care usage. Subjects and methods: A dataset consisting of 2542 weekly home care records for 1499 citizens (59% female, 41% male) with a mean age of 77 years (SD 10 years) was collected from a large municipality in Denmark. The data were recorded between January 1, 2021, and December 31, 2021. The dataset was divided into two cohorts. Subsequently, five machine learning-based survival analysis models were trained and evaluated using cross-validation. Results: The CoxBoost model showed the best discriminative performance with a mean 0.64 (95% CI 0.57–0.72) Harrell’s concordance index, indicating better ranking than chance-level by 14% on average. However, the model could not accurately predict when the next fall would take place. Conclusions: The proposed method enables professionals to assess individual fall risk by using home care records from an Electronic Health Record (EHR) system. This facilitates the initiation of targeted fall-prevention programs for those at highest risk. Additionally, it is expected that a risk-based approach can lead to a lower number needed to treat (NNT), indicating greater effectiveness of health interventions.
12 May 2023 in Research Square
Falling is the second leading cause of unintentional injury-related deaths, but fall injuries in older adults are often underreported, reported too late or reported with a lack of detail due to the stigma that surrounds falling. This makes it hard to locate and target the right citizens with fall prevention programs. In this paper, we propose a novel approach to assess fall risk over time and thus identify who would benefit from such a program. Using survival analysis, we train machine learning models to rank citizens by their predicted risk of getting a personal alarm within six months based on home care observations. Data were gathered from July 2021 to December 2021 for 57 types of home care from a large municipality in Denmark. After preprocessing, a total of 5713 observations for 346 citizens with mean age 81 (std. 5 years) are available for model training. Results show that a plain Cox proportional hazards model obtain a mean concordance index of 0.76 (95\% CI: 0.7-0.82), Random Survival Forest obtain 0.84 (95\% CI: 0.76-0.92) and XGBoost obtain 0.94 (95\% CI: 0.88-1) over 25 randomized folds using cross-validation. Censoring rate is 94.5\% on average. Most predictive features are help with personal hygiene, administration of medicine, excretions, cleaning and dispensing of medicine according to a SHAP summary plot. We also identify correlations between care usage in minutes per week and risk using SHAP dependency plots, but the correlation is not always positive. With our method, professionals at municipalities in Denmark or elsewhere, where home care services are recorded, will have a noninvasive decision-support tool that can predict individual fall risk over time.