Predictive Analytics in Healthcare: Real Examples from Home Health & Hospice

How Mosai Pulse Flags a Home Health Patient Before an ER Transfer

A home health patient with congestive heart failure is fourteen days into a sixty-day episode. Documented vitals were within an acceptable range. Her care plan is on schedule. Nothing in her most recent visit note would cause a clinical supervisor reviewing her chart to flag her for additional attention.

The predictive model sees something different. Evaluating her longitudinal record — her OASIS trajectory, vital sign trend across six weeks of prior episodes, weight movement over ten days, and clinician narrative notes processed through NLP — it identifies a trajectory pattern that has historically preceded acute deterioration. Her risk score moves into the elevated tier. She surfaces at the top of her supervisor's census view the next morning.

The supervisor reviews the flag, examines the clinical variables driving it, and assigns an unscheduled skilled nursing visit for that day. The visiting nurse finds early signs of fluid retention. A medication adjustment is coordinated with the attending physician. The patient does not go to the emergency department. The active care relationship is preserved, the remaining episode revenue is protected, and the readmission rate calculation for that quarter is not affected. Under a value-based care contract, the episode contributes positively to quality performance metrics rather than against them.

What Changed — and What Didn't

The model did not make a clinical decision. It identified a pattern and surfaced the patient. The supervisor exercised clinical judgment in reviewing the flag and assigning the visit. The nurse exercised clinical judgment in assessing the patient and recommending the intervention. The model created the intervention window. The clinicians used it.

What makes the reproducible:

Longitudinal clinical data at sufficient depth — OASIS patterns, vital sign trends, narrative documentation processed through NLP — allows a predictive model to catch a trajectory shift before it becomes a threshold event. Single-visit data cannot do this. Compassus, using Mosai Pulse across its home health operations, achieved a 10% reduction in hospitalization rate over two years — shifting from retrospective chart reviews to real-time, patient-level risk intelligence that enabled earlier, better-timed interventions.

How Mosai Transitions Surfaces a Patient at Elevated Mortality Risk

A home health patient with advanced heart failure and COPD is mid-episode. Her most recent skilled nursing visit was completed without documented alarm. From a chart-review perspective, she appears stable. What the chart review does not capture is her trajectory — the pattern forming across OASIS assessments, vital sign trends, functional decline indicators, and clinician narrative notes.

Mosai Transitions analyzes that full clinical picture against patterns historically associated with elevated mortality risk within approximately 90 days. Rather than relying on a single trigger, the model evaluates hundreds of clinical variables — trained on more than 460,000 patient records and continuously refreshed as new documentation is added. Her risk score moves into an elevated tier. The flag surfaces to her supervisor, who initiates a clinical review and opens a goals-of-care conversation with the patient and family — one that would not have happened based on the most recent visit note alone.

What Changed — and What Didn't

The Transitions model did not make a care planning decision. It identified a trajectory pattern historically associated with near-term mortality risk and surfaced it for clinical review. The supervisor decided what the flag warranted. The clinical team decided how to respond. The model created the information window. The clinicians used it.

What makes the reproducible:

Mosai Transitions was built specifically for home health and hospice: trained on more than 460,000 post-acute patient records, continuously refreshed, and designed to prioritize clinically meaningful longitudinal patterns over isolated findings — supporting earlier clinical review and goals-of-care conversations while there is still time to act.

How a Hospice Team Adjusts Care Intensity Based on a Muse Risk Score

A hospice patient with advanced cancer is three weeks into service. Her most recent skilled nursing visit documented manageable pain and a stable care plan — nothing that would trigger a rules-based alert. Mosai Muse, evaluating approximately 900 patient-specific data points — structured clinical features, NLP-processed visit narrative notes, and derived text-based features — identifies a trajectory shift. Her predicted likelihood of mortality within ten days moves into one of the model's higher-risk categories.

The hospice clinical supervisor reviews the flag and increases visit frequency, coordinates an interdisciplinary team discussion, and initiates a family conversation about what the patient's final days may look like. The right resources are in place before the patient enters the active dying phase — not after. The family has time to be present. The team has time to respond with intention.

What Changed — and What Didn't

Muse did not determine clinical status. It surfaced a trajectory shift warranting earlier attention. The supervisor decided to escalate. The team decided how to respond. The platform is designed to complement clinical judgment, not replace it.

What makes the reproducible:

Hospice mortality risk at this precision requires a model trained on hospice-specific populations at scale. Muse was developed on more than 4.5 million hospice patient visits and assigns patients to one of eight risk categories — giving supervisors more granular decision support than a binary classification. VIA Health Partners, using Mosai's predictive tools, achieved an 83% Hospice Visits in the Last Days of Life rate against a national average below 50%, with clinicians averaging more than 8.5 visits in the final seven days — evidence of what earlier risk identification produces at the agency level.

What Changes for a Clinical Supervisor Managing 40 Active Patients

Without predictive analytics, a clinical supervisor's morning triage is manual. She reviews recently documented patients, responds to overnight communications, checks in with field staff, and makes visit allocation decisions based on the information she has at hand — which is necessarily incomplete. Some patients who need attention today will not surface until their next scheduled visit.

With census-level predictive analytics — like those offered by Mosai Pulse for Home Health, Mosai Muse for Hospice, and Mosai Transitions for providers with both service lines — she opens a prioritized view of her forty active patients by current risk status. The three patients whose trajectory has shifted into the elevated risk tier are at the top of the list. She can see why — the clinical variables driving each flag are visible alongside the risk score. She assigns additional visits for two, initiates a care coordinator call for the third, and proceeds to manage the rest of her caseload from a position of clinical clarity rather than incomplete information.

What Changed — and What Didn't

The supervisor still made every clinical decision. She reviewed each flag, evaluated the rationale, and determined the appropriate response. What the model changed was the information state she was working from — and the time it took to achieve it. Manual triage across forty patients is a two-hour activity done rigorously. Census-level risk prioritization delivers the same information in a format that takes minutes to review. The time freed from triage is time available for the clinical work the triage was trying to enable.

What makes the reproducible:

Census-level risk prioritization requires that the model surface outputs at the population level — not just at the individual patient record level. Platforms that generate risk scores accessible only through individual chart navigation do not enable this workflow.

What These Examples Have in Common

Four different clinical scenarios, four different intervention points, the same underlying requirements. First, longitudinal clinical data at sufficient depth — each example requires a model reading a patient's trajectory across time, not a snapshot of a single visit. Second, post-acute-specific training data — the clinical variables that matter in home health and hospice are specific to this care setting, and a model that learned from general healthcare populations cannot reliably recognize risk in it. Third, census-level output that surfaces triage before a clinician has to go looking for it. Fourth, the human clinical judgment layer that acts on each signal. The agencies where these outcomes occur reliably are those where clinical workflows are designed to use predictive signals rather than around them.