AI Healthcare Follow-Up System Development

Industry: Healthcare / MedTech

Services: Conversational AI · Voicebot Development · Custom Web App · Clinical Automation

The Problem

Chronic disease management lives and dies on one thing: consistent follow-up. A doctor prescribes medication, sets a care plan, and sends the patient home, but what happens between visits is largely invisible.

A healthcare provider managing hundreds of chronic patients was facing a critical operational gap:

Follow-up calls were manual, inconsistent, and incomplete, and then nurses were overwhelmed

Doctors had no real-time visibility into how patients were doing between appointments

Medication non-compliance was going undetected until the next clinic visit, sometimes weeks later

Side effects and adverse reactions were being reported too late

Emergency situations, dangerous blood pressure spikes, and severe dehydration were missed entirely until they became crises.

Patient data was siloed; no unified health timeline existed per patient

Caregivers and family members had no structured way to stay informed

The question Cypherox was asked to solve: How do you give every chronic patient the experience of a dedicated clinical assistant at scale without adding a single staff member?

We designed and built a bot: a two-part intelligent healthcare platform combining a Kore.ai-powered AI Voice Bot that proactively calls patients on schedule, and the bot assistant, a web-based clinical command center for doctors and caregivers.

The system runs a fully closed loop: automated outbound call → structured health assessment → entity extraction → risk scoring → doctor dashboard → automated reporting with zero manual effort for routine cases.

Built on Kore. Ai's conversational AI platform, the bot conducts intelligent, adaptive health check-in calls with patients. It speaks naturally, handles unexpected responses gracefully, extracts clinical entities from speech, and escalates when it detects risk, all in real time.

Every call is structured across three sequential flows:

Outbound Call Initiation: The call is triggered automatically based on the patient's monitoring schedule stored in the database.

"Hello [Patient Name], my name is [Bot name], your doctor's assistant."

Small Talk Module (pilot phase) A brief, warm conversational opener based on the patient's known hobbies and interests, pulling data from the monitoring form (e.g., hiking, nature, travel) to establish rapport before the clinical questions begin.

"Could you please provide your patient ID number?"

Patient Says System Action Provides valid ID (e.g., 1485) Verified and fetches the patient profile, proceeds Provides unrecognized ID "Sorry, could not recognize your Patient ID, please try again." loops back Says, "I don't know my number." Routes to fallback retry → if still unresolved, escalates to a human agent

"Hello [Name], how are you doing overall compared to yesterday?"

Patient Says Bot Response "I feel very good." Logs positive sentiment, proceeds to Flow 2 "I feel about the same as yesterday." Logs neutral, proceeds "I feel bad, worse than yesterday." "I am sorry to hear that. What specifically makes you feel worse?" Something unrecognized "Sorry, I did not get that. Could you please repeat?" re-prompts up to 2x

If the patient describes feeling worse:

Clear answer (identifies symptom or reason) → logged with entity, proceeds

Unclear or no answer → "I am sorry, I did not catch that. Could you repeat the issue in different words?" → if still unclear, flags for human review

This is the bot's core clinical engine. Each area uses entity extraction, branching logic, and threshold-based escalation to go beyond simple yes/no answers, extracting clinically meaningful data from natural speech.

"Do you feel any regular pain today?"

Key Entities Extracted: Pain Rating (1–10) · pain_id (body part)

No / Not sure → logged, proceeds

Yes → "Where do you feel the pain?"

"Ok, noted. Any other pain?" → Yes loops back / No proceeds

"How is your hip pain compared to yesterday?"

A bit better / Same as yesterday / Hurting a lot today → all logged with comparative delta.

"Have you measured your blood pressure today?"

Key Entity Extracted: BloodPressure, two separate numeric values: systolic/diastolic

The bot is trained to recognize blood pressure stated in any natural format, "one twenty over eighty", "120/80", "it was high this morning", and normalize it to a numeric pair.

Reading Classification Action Systolic < 100 Low Flagged, doctor notified Systolic 101–130 Normal Logged, proceed Systolic 131+ High Emergency Alert Triggered

Patient hasn't measured → "Can you measure your blood pressure now? I will wait." or "That's okay, please try to measure it before tomorrow."

If a patient reports a very high reading → bot immediately escalates: "That reading is concerning, I am notifying your doctor right now." → Emergency node fires → doctor alerted in real time

"How did you sleep last night?"

Key Entities Extracted: SleepTime (hours) · Hours/quality descriptors (restful, broken, difficulty falling asleep)

Slept well → logged positive

Did not sleep well → "Tell me specifically what affected your sleep last night?"

"How many hours did you sleep?" → numeric extraction

Sleep quality index calculated and compared to the patient's baseline

Deteriorating patterns across multiple calls → flagged to doctor with commentary

"How much water/fluids have you had today?"

Key Entity Extracted: Litres_of_fluids

Patient states amount → Hydration Calculation Formula runs automatically (red node = critical threshold logic):

Critical low hydration: real-time notification pushed to doctor portal

"Have you taken today's dose of [PrescribedMedicine1]?"

Key Entities Extracted: Medication_name · Prescription_id · Prescription_admin (dosage, timing, method)

Patient Says Bot Action "Yes, I have" Compliance logged, proceed "What medication is that?" Bot reads full prescription: "You have [Med] prescribed to take 1 pill every morning with breakfast. Have you taken those?" "No, I have not." "I see, why haven't you taken the medicine?" → "I forgot." "Let me remind you, your doctor prescribed [Med] at [dosage/time]." Reminder logged → "It made me feel bad." "How did it make you feel bad? What did it do to you?" → Patient describes symptom "Alright, I will notify your doctor of this side effect." → Doctor alert triggered immediately → Unrecognized response "Sorry, I did not get that. Could you please repeat?"

Closing the loop, two-way communication between the patient and the doctor

Optional Announcement Module: If the doctor has left a message for the patient, the bot reads it at the end of the call:

"Your doctor has left the following message for you: [Doctor's message]."

Response Action "Yes" + says message Transcribed and logged → sent to doctor portal "Yes" (without message) "Ok, please say your message" → captures it "No" Proceeds to closing No answer / unclear "I am sorry, I did not catch that. Could you repeat the issue in different words?"

"Thank you for your time today. All the information will be sent to your doctor, and I will talk to you again on [Next Scheduled Date]."

Every patient response throughout the call is passed through Kore.ai's NLP sentiment engine. Beyond clinical data, the bot detects emotional indicators, distress, sadness, confusion, and anxiety in the patient's language and tone. A sentiment score is appended to every call record, feeding into the overall wellness assessment on the doctor's dashboard.

Built in React + Node.js + PHP, it is the web platform that receives everything the voicebot collects and transforms it into structured, actionable clinical intelligence.

Role Access Patient Caregiver Patient health summary, trend view, alerts Backend System Full data pipeline, call scheduling, automation Open Kore Assist AI agent assists during live calls Send Status Notification and reporting module

New patients are onboarded through a structured digital process:

Profile creation → Medical history → Current medications → Caregiver assignment

Hobbies and interests captured (used by voicebot for small talk personalization)

GDE Appointment Functionality for scheduling consultations

Onboarding call triggered → Onboarding form completed → Monitoring schedule auto-generated

Every call is visible in the Mamsistant dashboard in real time:

Open the Monitoring Form (auto-populating as the bot collects responses)

Open Kore AI Agent Assist (to monitor the live NLP session)

Fetch patient records mid-call

Route to Follow Monitoring Flow for escalated cases

End call manually if needed

Recording + call transcript stored automatically

Sentiment analysis score calculated and displayed

All data pushed to the MySQL database

Doctor portal updated in real time

Weekly and monthly automated email reports generated and dispatched

The structured post-call form gives doctors a complete, scannable health record for each call:

Assistant name · Attributed Patient · Time of Monitoring

Intro context (hobbies referenced by the bot during the call)

Loneliness Feeling score + Overall Wellbeing rating

Monitoring Q: "How are you doing overall today?"

Answer options + free-response commentary field

Doctor commentary column alongside patient response

Overall Wellbeing + Physical Wellbeing scores

Structured Q&A with answers and doctor commentary

All 5 monitoring area readings with trend indicators (↑ ↓ →)

Note field: "There will be one section of Mental Well-being", expandable per condition

When any monitoring area breaches a critical threshold, an emergency cascade fires:

Bot flags the node in real time (red node in flow)

Doctor receives instant push notification + email with patient name, the specific breach, and the exact response given.

Patient is immediately prompted for follow-up within the same call

Call transcript clipped to the relevant exchange and attached to the alert

Emergency entry logged in the patient file with a timestamp

Patients say "one twenty over eighty", "120/80", "my top number was high", all meaning the same thing. We trained custom Kore.ai entities to recognize every natural variation and normalize them into two clean numeric values before classification.

Five clinical areas, dozens of decision nodes, each with fallback handling, all stitched into one conversation that feels warm and human, not like a robotic questionnaire. This required careful dialog sequencing, session state management, and progressive disclosure of questions.

A missed hypertension spike or medication side effect is a patient safety failure. We built multi-signal escalation, a single critical reading fires immediately, but so does a pattern of borderline readings across multiple calls. The system catches both acute emergencies and slow deterioration.

Doctors need clinical precision with commentary fields and trend data. Caregivers need plain, emotional clarity. We served both from one database with two purpose-built portal views.

Elderly patients and those with cognitive difficulties often don't answer questions cleanly. Every single node has a structured fallback chain: first re-prompt, then rephrase, then flag for human review, so no patient is ever left stuck or frustrated.

"The bot calls our patients every week without fail. It catches things we would have missed for months. One patient's blood pressure had been creeping up across three consecutive calls; it flagged the pattern before it became a crisis. That kind of monitoring simply wasn't humanly possible before."