diff --git a/pyhealth/tasks/__init__.py b/pyhealth/tasks/__init__.py
index fb3c6966..16106363 100644
--- a/pyhealth/tasks/__init__.py
+++ b/pyhealth/tasks/__init__.py
@@ -62,3 +62,4 @@
     MutationPathogenicityPrediction,
     VariantClassificationClinVar,
 )
+from .dreamt_osa import DREAMTOSAClassification
diff --git a/pyhealth/tasks/dreamt_osa.py b/pyhealth/tasks/dreamt_osa.py
new file mode 100644
index 00000000..fd4d0af2
--- /dev/null
+++ b/pyhealth/tasks/dreamt_osa.py
@@ -0,0 +1,200 @@
+from typing import Any, Dict, List, Optional
+
+import polars as pl
+
+from .base_task import BaseTask
+
+
+class DREAMTOSAClassification(BaseTask):
+    """Patient-level OSA outcome tasks for the DREAMT dataset.
+
+    This task assumes you are using `DREAMTDataset`, which exposes patient-level
+    metadata such as:
+
+        - age
+        - gender
+        - bmi
+        - oahi
+        - ahi
+        - mean_sao2
+        - arousal_index
+        - medical_history
+        - sleep_disorders
+
+    We define OSA-related tasks using the AHI/OAHI values:
+
+    Tasks
+    -----
+    - "ahi_severity_4class"
+        Multi-class OSA severity from AHI:
+            0: AHI < 5        (normal)
+            1: 5 <= AHI < 15  (mild)
+            2: 15 <= AHI < 30 (moderate)
+            3: AHI >= 30      (severe)
+
+    - "ahi_binary_15"
+        Binary classification:
+            0: AHI < 15       (no / mild OSA)
+            1: AHI >= 15      (moderate / severe)
+
+    - "oahi_binary_5"
+        Binary classification:
+            0: OAHI < 5       (no sleep apnea by OAHI)
+            1: OAHI >= 5      (sleep apnea by OAHI)
+
+    Features
+    --------
+    By default, the feature vector for each patient is a dictionary of
+    clinical variables:
+
+        ["age", "gender", "bmi", "mean_sao2", "arousal_index",
+         "medical_history", "sleep_disorders"]
+
+    You can override this with `feature_keys` when instantiating the task.
+
+    Example
+    -------
+    >>> from pyhealth.datasets import DREAMTDataset
+    >>> from pyhealth.tasks import DREAMTOSAClassification
+    >>>
+    >>> dataset = DREAMTDataset(root="/path/to/dreamt/version")
+    >>> task = DREAMTOSAClassification(task="ahi_severity_4class")
+    >>>
+    >>> # later, in your dataloader construction:
+    >>> # samples = task(patient)  # where `patient` is from DREAMTDataset
+    """
+
+    # Registry of supported tasks
+    tasks = {
+        "patient_level": [
+            "ahi_severity_4class",
+            "ahi_binary_15",
+            "oahi_binary_5",
+        ]
+    }
+
+    def __init__(
+        self,
+        task: str,
+        feature_keys: Optional[List[str]] = None,
+    ) -> None:
+        if task not in self.tasks["patient_level"]:
+            raise ValueError(
+                f"Unsupported task '{task}'. "
+                f"Choose from: {self.tasks['patient_level']}"
+            )
+
+        self.task = task
+        self.task_name = f"DREAMTOSA/{task}"
+
+        # Default clinical features to include
+        self.feature_keys = feature_keys or [
+            "age",
+            "gender",
+            "bmi",
+            "mean_sao2",
+            "arousal_index",
+            "medical_history",
+            "sleep_disorders",
+        ]
+
+        # Patient-level tabular input
+        self.input_schema = {"feature": "tabular"}
+
+        # Label type depends on which task is chosen
+        if task == "ahi_severity_4class":
+            self.output_schema = {"label": "multiclass"}  # classes 0–3
+        elif task in {"ahi_binary_15", "oahi_binary_5"}:
+            self.output_schema = {"label": "binary"}
+        else:
+            # Should not happen because of the check above
+            raise ValueError(f"Unknown task: {task}")
+
+    def pre_filter(self, df: pl.LazyFrame) -> pl.LazyFrame:
+        """Optionally filter events before task construction.
+
+        For patient-level OSA tasks, we typically just pass everything through.
+        You can customize this to drop unrelated event types if needed.
+        """
+        return df
+
+    def __call__(self, patient: Any) -> List[Dict[str, Any]]:
+        """Build one or more samples for a single patient.
+
+        Each patient contributes exactly one sample for these tasks.
+        """
+        # Try to get split information if available, otherwise default to "train"
+        split = "train"
+        try:
+            split_events = patient.get_events("splits")
+            if len(split_events) == 1 and hasattr(split_events[0], "split"):
+                split = split_events[0].split
+        except Exception:
+            # If there is no "splits" table, we just keep the default.
+            pass
+
+        # Extract label depending on task
+        label = self._get_label_from_patient(patient)
+        if label is None:
+            # If we cannot compute a label (e.g., missing AHI/OAHI), skip this patient
+            return []
+
+        # Build feature dictionary from patient attributes
+        features: Dict[str, Any] = {}
+        for key in self.feature_keys:
+            value = getattr(patient, key, None)
+            features[key] = value
+
+        sample = {
+            "feature": features,
+            "label": label,
+            "split": split,
+            "patient_id": getattr(patient, "patient_id", None),
+        }
+        return [sample]
+
+    # ---------------------------------------------------------------------
+    # Internal helpers
+    # ---------------------------------------------------------------------
+
+    def _get_label_from_patient(self, patient: Any) -> Optional[int]:
+        """Compute the task-specific label from patient metadata."""
+        if self.task == "ahi_severity_4class":
+            ahi = getattr(patient, "ahi", None)
+            if ahi is None:
+                return None
+            try:
+                ahi_val = float(ahi)
+            except (TypeError, ValueError):
+                return None
+
+            if ahi_val < 5:
+                return 0  # normal
+            elif ahi_val < 15:
+                return 1  # mild
+            elif ahi_val < 30:
+                return 2  # moderate
+            else:
+                return 3  # severe
+
+        elif self.task == "ahi_binary_15":
+            ahi = getattr(patient, "ahi", None)
+            if ahi is None:
+                return None
+            try:
+                ahi_val = float(ahi)
+            except (TypeError, ValueError):
+                return None
+            return int(ahi_val >= 15.0)  # 1 = moderate/severe
+
+        elif self.task == "oahi_binary_5":
+            oahi = getattr(patient, "oahi", None)
+            if oahi is None:
+                return None
+            try:
+                oahi_val = float(oahi)
+            except (TypeError, ValueError):
+                return None
+            return int(oahi_val >= 5.0)
+
+        return None
diff --git a/tests/core/test_dreamt_osa.py b/tests/core/test_dreamt_osa.py
new file mode 100644
index 00000000..bff0fb56
--- /dev/null
+++ b/tests/core/test_dreamt_osa.py
@@ -0,0 +1,117 @@
+import numpy as np  # not strictly needed, but kept for consistency
+import polars as pl
+
+from pyhealth.data import Patient
+from pyhealth.tasks import DREAMTOSAClassification
+
+
+def test_dreamtosa_ahi_severity_4class():
+    # Minimal event-level DataFrame (no splits → default split="train")
+    df = pl.DataFrame(
+        {
+            "timestamp": [1],
+            "patient_id": ["P1"],
+            "visit_id": ["V1"],
+            "event_type": ["dreamt_sleep"],
+        }
+    )
+
+    # Create patient from raw event DataFrame
+    patient = Patient(
+        patient_id="P1",
+        data_source=df,
+    )
+
+    # Add DREAMT metadata attributes expected by the task
+    # (These are normally populated by DREAMTDataset / BaseDataset)
+    patient.ahi = 22.5  # moderate OSA → class 2
+    patient.age = 50
+    patient.gender = "F"
+    patient.bmi = 30.1
+    patient.mean_sao2 = 94.2
+    patient.arousal_index = 10.5
+    patient.medical_history = "HTN"
+    patient.sleep_disorders = "OSA"
+
+    # Initialize task
+    task = DREAMTOSAClassification(task="ahi_severity_4class")
+
+    # Run task
+    samples = task(patient)
+
+    # One sample per patient
+    assert len(samples) == 1
+    sample = samples[0]
+
+    # Check keys
+    assert "feature" in sample
+    assert "label" in sample
+    assert "split" in sample
+    assert "patient_id" in sample
+
+    # Check label mapping: AHI=22.5 → 2 (moderate)
+    assert sample["label"] == 2
+
+    # Check feature dict contents
+    features = sample["feature"]
+    assert isinstance(features, dict)
+    for key in [
+        "age",
+        "gender",
+        "bmi",
+        "mean_sao2",
+        "arousal_index",
+        "medical_history",
+        "sleep_disorders",
+    ]:
+        assert key in features
+
+    # With no splits table, default split is "train"
+    assert sample["split"] == "train"
+    assert sample["patient_id"] == "P1"
+
+
+def test_dreamtosa_ahi_binary_15_with_splits():
+    # Build event-level DataFrame including a "splits" event
+    df = pl.DataFrame(
+        {
+            "timestamp": [1, 2],
+            "patient_id": ["P2", "P2"],
+            "visit_id": ["V1", "V1"],
+            "event_type": ["dreamt_sleep", "splits"],
+            "split": [None, "val"],  # only used for the "splits" row
+        }
+    )
+
+    patient = Patient(
+        patient_id="P2",
+        data_source=df,
+    )
+
+    # Add DREAMT metadata
+    patient.ahi = 18.0  # >= 15 → label 1
+    patient.age = 60
+    patient.gender = "M"
+    patient.bmi = 28.0
+    patient.mean_sao2 = 93.0
+    patient.arousal_index = 15.0
+    patient.medical_history = "DM2"
+    patient.sleep_disorders = "OSA"
+
+    task = DREAMTOSAClassification(task="ahi_binary_15")
+    samples = task(patient)
+
+    assert len(samples) == 1
+    sample = samples[0]
+
+    # Binary label from AHI >= 15
+    assert sample["label"] == 1
+
+    # Split should be picked up from the "splits" event
+    assert sample["split"] == "val"
+    assert sample["patient_id"] == "P2"
+
+    features = sample["feature"]
+    assert isinstance(features, dict)
+    assert features["age"] == 60
+    assert features["gender"] == "M"