Learning from Few Subjects with Large Amounts of Voice Monitoring Data Jose Javier Gonzalez Ortiz John V. Guttag Robert E. Hillman Daryush D. Mehta Jarrad H. Van Stan Marzyeh Ghassemi Challenges of Many Medical Time Series • Few subjects and large amounts of data → Overfitting to subjects • No obvious mapping from signal to features → Feature engineering is labor intensive • Subject-level labels → In many cases, no good way of getting sample specific annotations 1 Jose Javier Gonzalez Ortiz Challenges of Many Medical Time Series • Few subjects and large amounts of data → Overfitting to subjects Unsupervised feature • No obvious mapping from signal to features extraction → Feature engineering is labor intensive • Subject-level labels Multiple → In many cases, no good way of getting sample Instance specific annotations Learning 2 Jose Javier Gonzalez Ortiz Learning Features • Segment signal into windows • Compute time-frequency representation • Unsupervised feature extraction Conv + BatchNorm + ReLU 128 x 64 Pooling Dense 128 x 64 Upsampling Sigmoid 3 Jose Javier Gonzalez Ortiz Classification Using Multiple Instance Learning • Logistic regression on learned features with subject labels RawRaw LogisticLogistic WaveformWaveform SpectrogramSpectrogram EncoderEncoder RegressionRegression PredictionPrediction •• %% Positive Positive Ours PerPer Window Window PerPer Subject Subject • Aggregate prediction using % positive windows per subject 4 Jose Javier Gonzalez Ortiz Application: Voice Monitoring Data • Voice disorders affect 7% of the US population • Data collected through neck placed accelerometer 1 week = ~4 billion samples ~100 5 Jose Javier Gonzalez Ortiz Results Previous work relied on expert designed features[1] AUC Accuracy Comparable performance Train 0.70 ± 0.05 0.71 ± 0.04 Expert LR without Test 0.68 ± 0.05 0.69 ± 0.04 task-specific feature Train 0.73 ± 0.06 0.72 ± 0.04 engineering! Ours Test 0.69 ± 0.07 0.70 ± 0.05 [1] Marzyeh Ghassemi et al. Learning to detect vocal hyperfunction from ambulatory neck-surface acceleration features: initial results for vocal fold nodules. IEEE Trans. Biomed6 Engineering Jose Javier Gonzalez Ortiz Summary • Our method learns features from large time series data • Reduces the need for laborious task-specific feature engineering • Applied to large voice monitoring dataset • Comparable performance to previous work that relied on expert engineered features 7 Jose Javier Gonzalez Ortiz.