Skip to content

progression

Jump to bottom
Yuwei (Evelyn) Zhang edited this page Mar 23, 2024 · 26 revisions

COX-based

Cox Survival model Youtube

  • combination of Kaplan miler and exponential model
  • assume that hazard changes over time, but hazard ratio is proportional
  • can only estimate hazard ratio of factors, cannot predict survival itself

Optimizing Survival Analysis of XGBoost for Ties to Predict Disease Progression of Breast Cancer

  • method: EXSA XGBoost + Cox, based on selected features
  • data: 12119 breast cancer patients
  • task: risk score, grouping

RNN/LSTM

T-LSTM

主要是EHR data, MIMIC-III

Recurrent neural networks for multivariate time series with missing values. Scientific reports 2018.

  • 增加了mask和interval信息，告诉模型missingness的信息，之前的工作只用了一个。
  • 进一步修改了GRU unit，加入了trainable decay
  • 思考：我们不管GRU，这些信息是不是也可以用来修改其他模型？比如feed给transformer之类的模型做为一个positional embedding/prompt？

ATTAIN: Attention-based Time-Aware LSTM Networks for Disease Progression Modeling IJCAI-19

  • 结合了之前分开有用过的attention和time-aware, added time interval to the attention

ODE

Neural ODE Neural Ordinary Differential Equations NeurIPS 2018 | YouTube

  • consider NN not as discrete layers but as a continuous function, a whole big block
  • not specifying depth of the network, let it learn by itself
  • better than RNN because: 1. more efficient 2. better at irregular time series
  • add ODE solver in the forward propagation

ODE2VAE [ODE2VAE: Deep generative second order ODEs with Bayesian neural networks] NeurIPS 2019

GRU-ODE-Bayes [GRU-ODE-Bayes: Continuous modeling of sporadically-observed time series] NeurIPS 2019 code

Integrating Expert ODEs into Neural ODEs: Pharmacology and Disease Progression NeurIPS 2021 | code

  • combine (explainable but limited) expert knowledge with ML latent representative power
image

Artificial intelligence-enabled detection and assessment of Parkinson’s disease using nocturnal breathing signals - nature medicine 2022

  • breath encoder: 8 residual blocks + 3 recurrent layers
  • PD encoder: attention layer to aggregate to global representation
  • classifier / severity predictor
  • used multitask learning in training, auxiliary task of predicting qEEG
  • domain invariant transfer learning: domain adversarial training of two sources in breath encoder, two discriminators for two classes
    • 还加了distribution calibration， ensemble of four calibrated models??
  • for severity predictor, added a consistency loss on predictions of different nights for the same subject.
  • 好复杂好多方法的叠加。。。

SSL-based

Unsupervised Learning of Disease Progression Models - KDD 2014

  • markov model, data of EHR on COPD patients

Longitudinal self-supervised learning

  • dataset: neuroimaging of Alzheimer (2641 of 811 patients) and alcohol dependence (1499 of 274+329)
  • Factor disentanglement: one dimension can be explained by a specific factor (brain age), invariant to other factors (e.g. gender, ethnicity)`
Screenshot 2023-10-27 at 18 08 13 * added a cosine loss while training auto-encoder, encourage the difference of two time steps to be in the certain direction * evaluated both frozen and fine-tuning * 思考：对于多个factor适用吗？对于并非单向的progression适用吗？

Modeling Disease Progression In Retinal OCTs With Longitudinal Self-Supervised Learning 19年一个workshop

  • dataset: 3308 scans from 221 patients
  • pretext task: predict the time interval between two scan images of the same patient
  • transfer task: intermediate state -> advanced stage ? by adding a classifier and fine-tune

Self-Supervised Pretraining and Transfer Learning Enable Flu and COVID-19 Predictions in Small Mobile Sensing Datasets Conference on Health, Inference, and Learning (CHIL) 2023 | code data on request

  • dataset: Homekit Flu Monitoring Study: 591k user-days of Fitbit data, 5196 participants, 6 months
  • pretext task: same user, reconstruct data, predict Domain Inspired Features (best!)
  • transfer task: infer positive / symptoms

Clone this wiki locally