Add MixLSTM model for temporally shifting clinical time-series (Oh et al. 2020)

docs/api/models.rst

@@ -174,6 +174,7 @@ API Reference
     models/pyhealth.models.MLP
     models/pyhealth.models.CNN
     models/pyhealth.models.RNN
+    models/pyhealth.models.mixlst
     models/pyhealth.models.GNN
     models/pyhealth.models.Transformer
     models/pyhealth.models.TransformersModel

docs/api/models/pyhealth.models.mixlstm.rst

@@ -0,0 +1,20 @@
+pyhealth.models.MixLSTM
+=======================
+
+The MixLSTM model from Oh et al. 2020, "Relaxed Parameter Sharing:
+Effectively Modeling Time-Varying Relationships in Clinical Time-Series"
+(https://arxiv.org/abs/1906.02898).
+
+MixLSTM addresses the problem of *temporal conditional shift* in clinical
+time-series, i.e., settings in which the relationship between input features
+and outcomes changes over the course of a patient's hospital stay. Instead
+of sharing a single set of LSTM parameters across all time steps, MixLSTM
+maintains ``K`` independent LSTM cells and, at every time step, computes a
+learned convex combination of their parameters using mixing coefficients.
+This enables smooth transitions between different temporal dynamics without
+requiring hard segment boundaries.
+
+.. autoclass:: pyhealth.models.MixLSTM
+   :members:
+   :undoc-members:
+   :show-inheritance: