🏥 Hospital Length of Stay (HLOS) Prediction

Deep Learning Pipeline for COVID-19 Inpatient Length of Stay Estimation

---

📋 Overview

This project builds an end-to-end deep learning pipeline for predicting Hospital Length of Stay (HLOS) for COVID-19 inpatients using clinical data from a Canadian hospital (508 patients). The pipeline covers data preprocessing, feature engineering, model training with hyperparameter tuning, and comprehensive evaluation across four distinct model architectures.

Models Implemented

#	Model	Input Data	Architecture
1	MLP	Static admission features	Feedforward network with BatchNorm + Dropout
2	LSTM	Sequential daily vitals/labs	Multi-layer LSTM with sequence-to-sequence output
3	Self-Attention	Sequential daily vitals/labs	Transformer-style multi-head self-attention
4	MLP + Word2Vec	Static features + NLP embeddings	MLP enriched with Word2Vec comorbidity embeddings

---

📊 Results

Model	MSE	RMSE	MAE	R²
MLP (static)	152.15	12.33	7.82	−0.20
LSTM (sequential)	103.17	10.16	6.80	0.13
Self-Attention (sequential)	57.37	7.57	4.33	0.52
MLP + Word2Vec	147.45	12.14	7.99	−0.16

Key Findings

• Self-Attention achieves the best performance (R² = 0.52), outperforming the LSTM by a significant margin. The attention mechanism can directly model inter-day dependencies without the information bottleneck of recurrent hidden states.
• Sequential models (LSTM, Attention) substantially outperform static models (MLP), confirming that the temporal trajectory of patient vitals and lab values carries critical prognostic signal beyond a single admission snapshot.
• Word2Vec comorbidity embeddings provide marginal improvement over the plain MLP (MSE 147 vs 152). The structured binary comorbidity columns already capture most of the signal, and the small corpus (404 patients) limits Word2Vec's representational power.

---

🗂️ Repository Structure

Hospital-Length-of-Stay-Prediction/
├── HLOS_Unified_Pipeline.py          # Full pipeline (standalone script)
├── notebooks/
│   └── HLOS_Unified_Pipeline.ipynb   # Jupyter/Colab notebook version
├── assets/
│   └── architecture_overview.png     # Model architecture diagram (optional)
├── requirements.txt                  # Python dependencies
├── .gitignore
├── LICENSE
└── README.md

---

🗃️ Dataset

Source: Canadian Hospital COVID-19 Inpatient Data
Format: Excel workbook with 4 sheets:

Sheet	Description	Shape
Data-at-admission	Static patient features at admission (demographics, vitals, labs, comorbidities)	508 × 55
Days-breakdown	Daily vitals and lab values for each patient (14-day records, 8 recorded days)	4,064 × 70
Hospital-length-of-stay	Target variable — actual length of stay + outcomes	508 × 11
Medication-Static-List	Reference list of all medications	31,254 × 2

Features Used

Static Features (58 total):

• Demographics: age, sex, height, weight
• Vitals: systolic/diastolic BP, heart rate, respiratory rate, SpO2, temperature, GCS (motor/verbal/eye)
• Lab Values: CBC (WBC, RBC, hemoglobin, hematocrit, MCV, MCH, MCHC, RDW, platelets), coagulation (APTT, PT, INR), liver (ALT, AST, bilirubin), renal (creatinine, sodium, potassium), inflammatory (lactate, CRP, ferritin, D-dimer, ESR), ABG (PaO2, PaO2/FiO2, pH)
• Clinical: intubation status, 18 binary comorbidity indicators
• NLP (Model 4 only): 50-dimensional Word2Vec embeddings from free-text comorbidity descriptions

Sequential Features (33 per timestep): Daily vitals + labs + intubation status across 8 recorded days.

Note: The dataset file (Canada_Hosp1_COVID_InpatientData.xlsx) is not included in this repository. Place it in the project root directory before running the pipeline.

---

🚀 Getting Started

Prerequisites

• Python 3.8+
• PyTorch 2.0+
• CUDA-capable GPU (optional but recommended)

Installation

git clone https://github.com/rishishanthan/Hospital-Length-of-Stay-Prediction.git
cd Hospital-Length-of-Stay-Prediction
pip install -r requirements.txt

Running the Pipeline

Option 1 — Python Script:

python HLOS_Unified_Pipeline.py

Option 2 — Jupyter Notebook:

jupyter notebook notebooks/HLOS_Unified_Pipeline.ipynb

Option 3 — Google Colab:

Click the "Open in Colab" badge at the top, upload the Excel dataset when prompted, and run all cells.

---

🏗️ Pipeline Architecture

┌─────────────────────────────────────────────────────────────────────┐
│                        RAW DATA (4 Excel Sheets)                    │
└──────────────┬──────────────────────────────────┬───────────────────┘
               │                                  │
       ┌───────▼───────┐                 ┌────────▼────────┐
       │ Data-at-       │                 │ Days-breakdown   │
       │ admission      │                 │ (sequential)     │
       └───────┬───────┘                 └────────┬────────┘
               │                                  │
    ┌──────────▼──────────┐            ┌──────────▼──────────┐
    │ Static Preprocessing │            │ Seq. Preprocessing  │
    │ • Merge with HLOS    │            │ • Merge with HLOS   │
    │ • Parse comorbidities│            │ • Compute remaining │
    │ • Encode categoricals│            │   LOS per day       │
    │ • Median imputation  │            │ • Median imputation │
    │ • StandardScaler     │            │ • 3D tensor build   │
    │ • Word2Vec (Model 4) │            │ • StandardScaler    │
    └──────┬──────┬───────┘            └──────┬──────┬──────┘
           │      │                           │      │
     ┌─────▼──┐ ┌─▼──────────┐       ┌───────▼──┐ ┌─▼───────────┐
     │ MLP    │ │ MLP+Word2Vec│       │  LSTM    │ │ Attention   │
     │Model 1 │ │  Model 4   │       │ Model 2  │ │  Model 3    │
     └────┬───┘ └─────┬──────┘       └────┬─────┘ └──────┬──────┘
          │           │                    │              │
          └─────┬─────┘                    └──────┬───────┘
                │                                 │
         ┌──────▼─────────────────────────────────▼──────┐
         │           Evaluation & Comparison              │
         │  MSE · RMSE · MAE · R² · 5-Trial Analysis     │
         └────────────────────────────────────────────────┘

---

⚙️ Hyperparameter Configurations Tested

MLP Configurations

Config	Hidden Layers	Dropout	Learning Rate	Test MSE
1 ✅	[128, 64]	0.3	1e-3	152.15
2	[256, 128, 64]	0.3	1e-3	170.97
3	[128, 64, 32]	0.2	5e-4	190.86
4	[256, 128, 64, 32]	0.3	1e-3	165.23

LSTM Configurations

Config	Hidden Dim	Layers	Dropout	Learning Rate	Test MSE
1	64	1	0.0	1e-3	108.58
2	64	2	0.2	1e-3	104.00
3	128	2	0.2	5e-4	108.55
4 ✅	128	3	0.3	5e-4	103.17

Self-Attention Configurations

Config	d_model	Heads	Layers	Dropout	Learning Rate	Test MSE
1	64	4	2	0.2	1e-3	61.95
2 ✅	64	4	3	0.2	5e-4	57.37
3	128	4	2	0.2	5e-4	70.21
4	128	8	3	0.3	5e-4	64.08

MLP + Word2Vec Configurations

Config	Hidden Layers	Dropout	Learning Rate	Test MSE
1	[128, 64]	0.3	1e-3	149.05
2 ✅	[256, 128, 64]	0.3	1e-3	147.45
3	[256, 128, 64, 32]	0.2	5e-4	181.87

---

📈 Multi-Trial Analysis (5 Trials, Random 80% Training Subsets)

Model	Training MSE (mean ± std)	Test MSE (mean ± std)
MLP	35.70 ± 19.05	159.47 ± 24.24
LSTM	122.58 ± 23.10	115.96 ± 8.26

The LSTM shows lower variance across trials (std = 8.26 vs 24.24), indicating more stable generalization. The MLP's large train-test gap confirms overfitting on the small static dataset.

---

💡 Potential Improvements

• Imputation: KNN or iterative imputation instead of simple median
• Target transform: Log-transform the right-skewed HLOS distribution
• Architecture: Bidirectional LSTM, GRU variants, Temporal Fusion Transformers
• NLP: Pre-train Word2Vec on a larger medical corpus (e.g., PubMed/MIMIC-III notes)
• Ensembling: Combine predictions from multiple models
• Feature engineering: Add missingness indicator features for highly sparse columns (PaO2, ferritin, ESR have >90% missing)

---

🛠️ Tech Stack

• Deep Learning: PyTorch
• NLP: Gensim (Word2Vec)
• Data Processing: Pandas, NumPy, Scikit-learn
• Environment: Google Colab / Jupyter Notebook

---

📄 License

This project is licensed under the MIT License — see the LICENSE file for details.

---

👤 Author

Rishi Shanthan Bhagavatham
MS in Engineering Science (Data Science) — University at Buffalo (SUNY)