From: DeepMPM: a mortality risk prediction model using longitudinal EHR data
Notation | Meaning |
---|---|
D | Diagnoses codes set, \(D=\{d_{1}, d_{2}, \ldots , d_{k}\}\) |
L | DRGs codes set, \(L=\{l_{1}, l_{2}, \ldots , l_{s}\}\) |
\(X_t\) | Representation vector of diagnosis |
\(P_t\) | Representation vector of treatment |
\(x_t\) | Diagnoses codes of a record, \(x_t \in \{ 0,1\}^{|D|}\) |
\(p_t\) | DRGs codes of a record, \(p_t \in \{ 0,1\}^{|L|}\) |
\(W_{xemb}\) | Weight of embedding layer for diagnoses codes |
\(W_{pemb}\) | Weight of embedding layer for DRGs codes |
\(f_t\) | Forget gate of LSTM at time step t |
\(W_f\) | Weight of the forget gate of LSTM |
\(i_t\) | Input gate of LSTM at time step t |
\(W_i\) | Weight of the input gate of LSTM |
\(\tilde{C_t}\) | Candidate cell state of LSTM at time step t |
\(C_t\) | Cell state of LSTM at time step t |
\(o_t\) | Output gate of LSTM at time step t |
\(W_o\) | Weight of the output gate of LSTM |
\(h_t\) | Hidden state of LSTM at time step t |
\(m_t\) | Type of hospitalization |
\(q_t\) | Hospital stay vector |
\(U_i\) | Weight of \(h_{t-1}\) in the input gate of Care-LSTM |
\(U_f\) | Weight of \(h_{t-1}\) in the forget gate of Care-LSTM |
\(P_f\) | Weight of \(P_{t-1}\) |
\(Q_f\) | Weight of \(q_{\Delta _{t-1:t}}\) |
\(q_{\Delta _{t-1:t}}\) | Hospital stay during \(\Delta _{t-1:t}\) |
\(\Delta _{t-1:t}\) | Adjacent hospital stay intervals |
\(U_o\) | Weight of \(h_{t-1}\) in the output gate of Care-LSTM |
\(P_o\) | Weight of \(P_t\) |
\(g_t\) | Output of the hidden layer of Care-LSTM at time step t |
\(\alpha _t\) | Variable-level weight vector, \(\alpha _t \in [0,1]\) |
\(W_\alpha ^T\) | Weight matrix in attention module |
\(e_t\) | Output of the hidden layer of Care-LSTM at time step t |
\(\beta _t\) | Visit-level weight vector, \(\beta _t\in [-1,1]\) |
\(W_\beta\) | Weight matrix in attention module |
\(r_t\) | Harmonic weight coefficient |
\(w_t\) | Final weight vector of the two-level attention module |
\({\bar{h}}\) | Patient health status vector |