<!-- class: center, middle --> class: center, middle name:opening ## Learning a Data-Driven Nosology: ## Progress, Challenges & Opportunities <!-- <img src="images/psych.jpg" STYLE="position:absolute; TOP:200px; LEFT:110px; WIDTH:700px;"/> --> --- ## Steps <br><br> 1. Acquire data 2. Build and run pipelines 3. Develop and apply clustering methods --- ### Progress: Datasets Acquired <img src="images/m2gtab.png" STYLE="position:absolute; TOP:200px; LEFT:00px; WIDTH:900px;"/> --- ### Progress: Pipeline Built! <img src="images/ndmg-pipeline.png" STYLE="position:absolute; TOP:200px; LEFT:0px; WIDTH:900px;"/> --- class: center, middle ## Clustering Methods: In Progress -- 1) Don't find spurious clusters <br> 2) Do find legitimate clusters --- ### Challenge #1: Spurious Clusters <img src="images/spikes.png" STYLE="position:absolute; TOP:150px; LEFT:150px; HEIGHT:500px;"/> --- ### Challenge #1: Spurious Clusters <img src="images/spikes_kde.png" STYLE="position:absolute; TOP:150px; LEFT:150px; HEIGHT:500px;"/> --- ### Challenge #1: Spurious Clusters <img src="images/spikes_bic1.png" STYLE="position:absolute; TOP:150px; LEFT:150px; HEIGHT:500px;"/> --- ### Challenge #1: Spurious Clusters <img src="images/spikes_bic2.png" STYLE="position:absolute; TOP:150px; LEFT:150px; HEIGHT:500px;"/> --- ### Challenge #1: Spurious Clusters <img src="images/spikes_bic3.png" STYLE="position:absolute; TOP:150px; LEFT:150px; HEIGHT:500px;"/> --- ### Challenge #1: Spurious Clusters <img src="images/spikes_bic4.png" STYLE="position:absolute; TOP:150px; LEFT:150px; HEIGHT:500px;"/> --- ### Challenge #1: Spurious Clusters <img src="images/spikes_bic5.png" STYLE="position:absolute; TOP:150px; LEFT:150px; HEIGHT:500px;"/> --- ### Challenge #1: Spurious Clusters <img src="images/spikes_bic1000.png" STYLE="position:absolute; TOP:150px; LEFT:50px; WIDTH:800px;"/> <!-- --- ### Challenge #1: Spurious Clusters real data goes here --> --- class:center, middle ### Opportunity #1: Spurious Clusters <br> Standard tools find spurious clusters <br> Fancy tools require too much data to be practical <br> .w[**Gap**: Clustering tools with appropriate statistical guarantees] --- ### Challenge #2: Legitimate Clusters <img src="images/batch.png" STYLE="position:absolute; TOP:150px; LEFT:50px; WIDTH:800px; HEIGHT:400px"/> --- ### Challenge #2: Legitimate Clusters <img src="images/batch_sex.png" STYLE="position:absolute; TOP:150px; LEFT:50px; WIDTH:800px; HEIGHT:400px"/> --- class: center, middle ### Opportunity #2: Legitimate Clusters <br> Thus far, harmonzing processing does not remove batch effect <br> Can further harmonizing processing mitigate batch effect? <br> If not, can harmonizing acquisition mitigate batch effect? <br> .w[**Gap**: Pipelines that harmonize across datasets, and data to evaluate pipelines] --- ### Future Outlook - Microarrays never mitigated batch effects sufficiently - Neither microarrays nor MRI "count" - fMRI has another degree of freedom: stimulus -- ### Much More To Do! <img src="images/lion2.JPG" STYLE="position:absolute; TOP:350px; LEFT:0px; WIDTH:400px;"/> <img src="images/world.gif" STYLE="position:absolute; TOP:350px; LEFT:400px; WIDTH:500px;"/> <DIV style="position: absolute; top:550px; left:120px; ">(i love this guy)</DIV> <DIV style="position: absolute; top:620px; left:340px; "><a href="mailto:jovo@jhu.edu">jovo@jhu.edu</a></DIV> --- #### Challenge #3: Overlapping Clusters <img src="images/overlap.png" STYLE="position:absolute; TOP:150px; LEFT:150px; HEIGHT:500px"/> --- #### Challenge #3: Overlapping Clusters <img src="images/overlap.png" STYLE="position:absolute; TOP:150px; LEFT:50px; WIDTH:800px; HEIGHT:50px"/> <img src="images/levels0.png" STYLE="position:absolute; TOP:200px; LEFT:50px; WIDTH:800px;"/> --- #### Challenge #3: Overlapping Clusters <img src="images/overlap.png" STYLE="position:absolute; TOP:150px; LEFT:50px; WIDTH:800px; HEIGHT:50px"/> <img src="images/levels1.png" STYLE="position:absolute; TOP:200px; LEFT:50px; WIDTH:800px;"/> --- #### Challenge #3: Overlapping Clusters <img src="images/overlap.png" STYLE="position:absolute; TOP:150px; LEFT:50px; WIDTH:800px; HEIGHT:50px"/> <img src="images/levels2.png" STYLE="position:absolute; TOP:200px; LEFT:50px; WIDTH:800px;"/> --- #### Challenge #3: Overlapping Clusters <img src="images/overlap.png" STYLE="position:absolute; TOP:150px; LEFT:50px; WIDTH:800px; HEIGHT:50px"/> <img src="images/levels3.png" STYLE="position:absolute; TOP:200px; LEFT:50px; WIDTH:800px;"/> --- #### Challenge #3: Overlapping Clusters <img src="images/overlap.png" STYLE="position:absolute; TOP:150px; LEFT:50px; WIDTH:800px; HEIGHT:50px"/> <img src="images/levels4.png" STYLE="position:absolute; TOP:200px; LEFT:50px; WIDTH:800px;"/> --- #### Challenge #3: Overlapping Clusters <img src="images/overlap.png" STYLE="position:absolute; TOP:150px; LEFT:50px; WIDTH:800px; HEIGHT:50px"/> <img src="images/levels5.png" STYLE="position:absolute; TOP:200px; LEFT:50px; WIDTH:800px;"/> --- class: center, middle ### Opportunity #3: Overlapping Clusters <br> Any tree with thresholds will put arbitrarily close individuals arbitrarily far away <br> Trees that put individuals near boundary in multiple groups could mitigate this issue <br> .w[**Gap**: Tree learning tools with appropriate statistical guarantees]