README_old_Project_process.md

Parkinson Disease Project - DSR

Subject: Modelling Parkinson Disease Progression and predicting best treatment strategies for PD patients with Machine Learning/Deep Learning

AI & Healthcare topics: diagnosis prediction/accuracy, precision medecine with personalized treatments

0. Project context

• number of hours to work on it: ~350-400h
• Project output:
• Github repo with reproducible code
• Oral Presentation (~10/15min + questions) in front of panel of companies and data scientists in Berlin
• Slides Deck
• Optional: Video about the Project Process, Medium Article, Scientific paper (if relevant and possible)

1. Abstract / Project objectives:

The main goal of this project is predicting Parkinson Disease (PD) Progression over time for a set of patients from a clinical study.

More precisely, the project objectives are: Starting from a dataset of Electronic Health Records of patients with 4 different clinical diagnosis (PD diagnosed patients, 400 de novo PD subjects - newly diagnosed and unmedicated, SWEDD and healthy control subjects),I will:

• Classify clinical diagnosis for each patient
• Predict Parkinson Disease Progression over time using:
  • 2a. First, only biological biomarkers .
  • 2b. Add in a second time MRI data (raw brain scans from the PPMI dataset) . (the idea is to compare the performance of the models with and without MRI data, as performing MRI is expensive for hospitals and also not easy to process as data).
• Predict best treatment strategy associated to each patient and each future disease state
• If enough time: predicting patients' time of death (see recent article about Google having done such predictions)
• If enough time: ANALYZE TEMPORAL PATTERNS OF BIOMARKERS USING LONGITUDINAL STABILITY SELECTION (as done on AD in Fused Sparse Group Lasso Paper)

NB: SWEDD= Scans without evidence of dopamine deficit: patients who look like they have PD in terms of symptom but subsequent functional imaging assessment does not confirm this. source: http://www.acnr.co.uk/SO10/ACNRSO10_30_SWEDD_article.pdf

2. MAIN DATASET:

http://www.ppmi-info.org/access-data-specimens/download-data/

Data Pop : 2144 patients (as of SCREEN tables). 1052 PD people (PD (Gen and non Gen cohort), SWEDD, PRODROMA)

Patients Hospital visits timeline (timeseries period):

• 15 visits besides screening, Baseline and unscheduled visits.
• -> V04b: every 3 month / V05b -> V13: every 6 months / V14, V15 and beyond; every 12 months.

Schedule of activities for the clinical studies here:

(http://www.ppmi-info.org/wp-content/uploads/2018/02/PPMI-AM-13-Protocol_SCHEDULE-OF-ACTIVITIES-1.pdf)

data dictionnary annotated here:

https://docs.google.com/spreadsheets/d/1Q-0zAG_oBfuo21s5xzN6Vwck5NT0GyP0K3KxiWpxO58/edit#gid=782043355 (work in progress)

• Complete dataset can be downloaded here: https://github.com/AMDonati/parkinson-disease-project/blob/master/data/PPMI-final-dataset-382018.zip
• additionnal dataset info: https://www.ppmi-info.org/study-design/research-documents-and-sops/

3. DS main challenges for this project:

• Multi-modal (dataframes + MRI scans) multi-label classification problem of timeseries with features selection depending on the label prediction (the PD predominant biomarkers are depending on the disease state)

4. Use cases:

• FOR DOCTORS:
  • Early detection of PD subjects
  • Improving disease accuracy,
  • Anticipation of their patients' disease evolution
  • support to decision for choosing treatments for patients
• FOR PD PATIENTS:
  • For people with risk of getting PD: early detection and monitoring
  • For early-stage/late-stage PD people: monitoring better their disease by knowing the expected evolution
  • Finding the best treatment that will improve the quality of life of PD people
• FOR HOSPITALS:
• cost reductions by improved patient management & treatment optimization
• operations improvement by predicting patients future visits, med supplies, etc...

---

5.DS Process for the project

1. Data Processing/Cleaning: Merge/Join of Tables (eventually using PyTables)

2. EDA: timeseries analysis & plotting

look @ sparsity, irregularity,trends...

3.Feature Engineering: (if classic ML is used)

• Manual Feature Engineering with the help of doctors
• Statistical techniques to find correlation between features
• Classic Feature Engineering Techniques for ML -> use manifold learning:
• PCA: http://scikit-learn.org/stable/modules/generated/sklearn.decomposition.PCA.html
• t-SNE: https://lvdmaaten.github.io/tsne/ https://bl.ocks.org/Fil/b07d09162377827f1b3e266c43de6d2a
• isomap: http://scikit-learn.org/stable/modules/manifold.html . https://github.com/cc-skuehn/Manifold_Learning . https://towardsdatascience.com/reducing-dimensionality-from-dimensionality-reduction-techniques-f658aec24dfe
• autoencoders

4. ML/DL models to try:

• Baseline: a statistical model > the one for a multivariate timeseries with endogeneous variables without trends and without seasonality = VARMA.
  • other possibilities: discounted median, persistence model.

Some references: https://machinelearningmastery.com/persistence-time-series-forecasting-with-python/

• "Explainable" models: gradient-boosted trees, Random Forest

• DL Models to be modified to include incertainty with bayesian models

  • Hidden Markov Models

  About HMMs for disease progression: https://web.engr.oregonstate.edu/~lif/nips2015_CTHMM_learning_camera_ready.pdf About HMMs for TS classification: https://towardsdatascience.com/hidden-markov-models-for-time-series-classification-basic-overview-a59b74e5e65b

  • LSTM, GRU

  about pro/cons of LSTM vs GRU: https://datascience.stackexchange.com/questions/14581/when-to-use-gru-over-lstm
  A review of GRU for sequence modeling: https://arxiv.org/pdf/1412.3555v1.pdf

  • Seq2seq, 1D-CNN, MFCN LSTM?

reference paper for timeseries foreacast of EHR: https://arxiv.org/pdf/1706.03446.pdf (to print)

About chossing between HMM & RNNs:

• https://www.quora.com/When-would-one-use-a-Hidden-Markov-Model-instead-of-a-Recurrent-Neural-Network

• https://stats.stackexchange.com/questions/20429/what-are-the-differences-between-hidden-markov-models-and-neural-networks

• https://pub.uni-bielefeld.de/download/2903474/2907910

  NB: maybe the problem of allocating treatments to disease states over time could be solved by reinforcement learning as well

  See these links:

• [Google scholar search for RL for adaptative treatments strategies] (https://scholar.google.com/scholar?oi=gsb95&q=reinforcement%20learning%20for%20medical%20treatments%20decisions&lookup=0&hl=en)
• [An experimental design for the development of adaptive treatment strategies] (https://deepblue.lib.umich.edu/bitstream/handle/2027.42/39201/2022_ftp.pdf?sequence=1&isAllowed=y)
• [Informing sequential clinical decision-making through reinforcement learning: an empirical study] (https://link.springer.com/content/pdf/10.1007/s10994-010-5229-0.pdf)
• [Customizing Treatment to the Patient: Adaptive Treatment Strategies] (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC1924645/)

• see also this article from MIT news using RL for reducing drug doses: http://news.mit.edu/2018/artificial-intelligence-model-learns-patient-data-cancer-treatment-less-toxic-0810

6. Outputs of the model . For each patient:

• a vector of disease states for different timestamps (multiple labels for each timestamp corresponding to UPDRS scale) .

link to UPDRS scale: UPDRS: http://www.etas.ee/wp-content/uploads/2013/10/updrs.pdf

• The associated treatment strategy related to the disease state (multiple labels as well)
• The estimated time of death

7. Data Visualisation to present the results . Plots comparating the performance of the different models for different metrics

4. DS tools to be used (Python Libraries...):

DS Process	Tool/Python lib	Comments
0. Project Management	Trello, slack, Github
1. Data Processing	PyTables? or only pandas	Mosty done
2. EDA	Cufflinks/plotly, Seaborn, Scipy, statsmodels.tsa	Mostly done
3. Feature Engineering	sci-kit learn	Manually done for now .
4a. Classic ML	sci-kit learn	probably no time for that
4b. DL	keras with TensorFlow BE, Tensorboard, Google Open Cloud/AWS	tutos started
5. Dataviz	TBD, probably Seaborn

Links:

• http://www.statsmodels.org/stable/tsa.html
• for quick dataviz on dataframes: use cufflinks: https://plot.ly/ipython-notebooks/cufflinks/
• For DL:
  • [NN Playground] (https://playground.tensorflow.org/#activation=tanh&batchSize=30&dataset=spiral&regDataset=reg-plane&learningRate=0.3&regularizationRate=0&noise=0&networkShape=4,2&seed=0.51575&showTestData=false&discretize=false&percTrainData=50&x=true&y=true&xTimesY=false&xSquared=false&ySquared=false&cosX=false&sinX=false&cosY=false&sinY=false&collectStats=false&problem=classification&initZero=false&hideText=false)
  • Tensorboard
  • GPUs: Google Open Cloud (Free), AWS

Python tricks to used

• Pickles

https://docs.python.org/2/library/pickle.html . https://pythontips.com/2013/08/02/what-is-pickle-in-python/

---

8. Useful tutos to do for the project

Deep Learning - General

• http://course.fast.ai/index.html (RNNs, CNNs, autoencoders...)
• DL models hyperparameters tuning: https://www.datacamp.com/courses/deep-learning-in-python?utm_source=fb_paid&utm_medium=cpc&utm_campaign=dc_registered_visitors
• List of NN architectures description: http://www.asimovinstitute.org/neural-network-zoo/ http://www.asimovinstitute.org/blog/
• Chapter advanced use of Keras from F.Chollet book.

Biomedical images processing

• Deep Learning for Heathcare Image Analysis: https://nvidia.qwiklab.com/quests/14
• Deep Learning for Medical Imaging: https://www.youtube.com/watch?v=2_Jv11VpOF4
• Tuto to read fMRI data: https://medium.com/coinmonks/visualizing-brain-imaging-data-fmri-with-python-e1d0358d9dba (see associated github repo)
• THE BEST: https://pyscience.wordpress.com/2014/09/08/dicom-in-python-importing-medical-image-data-into-numpy-with-pydicom-and-vtk/

9. List of contacts that could help me & summary of discussions (other than mentor(s))

see here: https://github.com/AMDonati/parkinson-disease-project/wiki/useful-contacts-&-discussions-summary

10. Reference Documents

• see list here: https://github.com/AMDonati/parkinson-disease-project/wiki/reference-documents
• See papers summary here: https://docs.google.com/spreadsheets/d/1TEL4umtQPdPPhzf092iMQgvmi9pID5UMrCdndo6ZJsE/edit#gid=0

11. Additonal datasets that could be used:

See here: https://github.com/AMDonati/parkinson-disease-project/wiki/Other-PD-medical-datasets

---

Parkinson News to check:

https://parkinsonsnewstoday.com/2018/07/31/ema-endorses-use-of-imaging-radioactive-tracer-to-improve-parkinsons-clinical-trial-recruitment/