Open-source IV drip flow monitor for pediatric humanitarian care.

Pediatric IV delivery is silently dangerous in low-resource settings.

Gravity-fed IV drips are the default in humanitarian and rural healthcare. Without an infusion pump, flow rate is set by hand-counted drops — and a single missed adjustment can deliver a fatal volume to a child in under an hour.

Dual-beam optical drop detection.

Two IR beams cross the drop path inside the drip chamber, vertically separated by 10.0 mm. Each drop yields two independent measurements per crossing: the shadow duration τ — how long the drop occludes a beam — sets the drop's diameter, and the inter-beam transit Δt sets its velocity. Volume follows from a sphere model; flow rate is the product of per-drop volume and drip cadence — independent of the drip-set's nominal gtt·mL⁻¹ rating.

Board architecture.

A 50 × 45 mm 2-layer board, hand-solderable for low-volume humanitarian assembly. Back side (STM-side) shown. Click any numbered callout for the schematic snippet, the rationale, and the ADR that locked the decision.

3D render Routing

True size

50 × 45 mm · 2-layer · F.Cu shown · 3× display scale

1234567

click me!

Hotspot 1 · MCU

STM32G071C8T6 · core + decoupling

[ KiCad schematic snippet ]

48-pin LQFP, single 3.3 V rail with 4× 100 nF decoupling. Migrated from STM32G030 → G031 → G071 because G030/G031 lack COMP1 and the DAC required for the threshold reference. Pin-compatible swap, no layout cost.

↪ Architecture decisions

Single Docker command regenerates every figure in this page from the raw CSVs (Zenodo deposit pending mint). Hardware files are KiCad 8 native; firmware builds on STM32CubeIDE 1.16; enclosure prints in ASA on a 0.4 mm nozzle.

Predicted accuracy from first principles, before any bench data.

A Monte Carlo error budget propagates measurement uncertainty through the firmware's volume-inversion math to predict per-drop MAPE, Bland-Altman Limits of Agreement, and calibration convergence — committed to git before the validation campaign began. The four headline numbers below are the pre-bench prediction; the Limitations section carries the measured cross-session Bland-Altman so the predict-vs-measure gap is reviewable side-by-side. Source notebook: analysis/notebooks/error_budget.ipynb.

These four numbers were the pre-bench prediction. The MC budget correctly identified print tolerance as the dominant predicted source (1.54 percentage points of the 3.13 % MAPE) but under-estimated the residual ~6×; the gap turns out to be the position-dependent contamination of the chord measurement, visible directly in the raw photodiode waveforms below. The capture firmware (ENABLE_RAW_CAPTURE on the feat/raw-capture-rev-c branch) is not the Rev-B submission firmware; submission firmware = HEAD of v2. See §17 in Limitations for the full architectural treatment.

cd analysis
docker compose up regenerate-error-budget

cd analysis
docker compose up regenerate-figures-sample   # cold-clone test (sample data)
docker compose up regenerate-figures          # full campaign data

What Rev-B does not resolve.

Six load-bearing limitations of the Rev-B validation, in order of impact on the conclusions. The full list of 17 is in docs/limitations.md. Each item below links to its full section. The position-dependence finding (§17) is the single most important Rev-C-relevant learning of the bench campaign.

1. §17

   Position-dependence — V_CAL_K does not generalise across mount heights.

   Architectural

   On the afternoon of 2026-05-13 the device was mounted at four mount positions on the same drip chamber (same set, same fluid, same board). For the shipped mean-of-pulses algorithm, K_fit varied from 0.42 to 1.27 across positions (3× spread); across all eight algorithm variants trialled, K_fit spanned 0.42 to 1.98 (4.7× spread driven by mount geometry alone). Three contaminants of the chord measurement were isolated: an umbilical at TOP when mounted close to the drip orifice, backsplash at BOT when mounted close to the pool, and a threshold-vs-peak mismatch that shifts with the boot-cal baseline. Several algorithm variants (BOT_core only, TOP_low only, adaptive EMA, hybrid splash-detection) were tried; no single (algorithm, K) combination passed ±5 % at all positions. Conclusion: with Rev-B optics, the optical chord-time architecture does not deliver a position-invariant V_CAL_K. The two beams do deliver position-agnostic velocity and reliable drop counting — what they don't deliver is a position-invariant chord. The Rev-C path is named: a second pair of horizontally-separated beams, wider TOP-to-BOT separation, and TOP/BOT gain equalisation.

2. §1

   Drops oscillate between oblate and prolate after detachment.

   Fundamental

   The chord-time architecture measures each drop's vertical extent and assumes a sphere. After detachment, drops oscillate between oblate and prolate as surface tension pulls them toward equilibrium (Rayleigh 1879); the beam catches each drop at an unpredictable phase. The run-mean bias averages out (V_CAL_K = 1.27 works on the run mean) but the per-run residual stays at ±30 %. A Rev-C front-end with horizontally-separated beam pairs would let firmware infer the oscillation phase.

3. §2

   Two-orifice divergence invalidates per-drop matching.

   Methodological

   The device measures drops at the drip-chamber orifice; the scale captures mass at the end-of-tubing orifice. These two orifices have different geometry and surface-tension dynamics, so they drip at different rates with different drop sizes. The 2026-05-13 campaign saw end-orifice drop counts 1.5–2× the chamber-drop counts. Mass conservation still gives a clean shared quantity (total mass passed), so the valid ground truth at the per-drop level is V_true_chamber = mass / N_UART, not a per-drop pairing. tools/match_scale_drops.py is preserved as an exploratory diagnostic only.

4. §3

   TOP-vs-BOT optical asymmetry (backsplash + intrinsic).

   Optical

   The TOP and BOT photodiodes report systematically different pulse widths for the same drop. The 2026-05-13 backsplash falsifiability test repositioned the device higher on the chamber; the BOT/TOP ratio fell from 1.75–2.50× to ~1.13×, suggesting the BOT photodiode sees a real splash-driven extension that the threshold-margin clips. A Rev-C optical front-end with the BOT beam moved further from the pool, plus TOP/BOT gain equalisation, would close the gap.

5. §5

   V_CAL_K is fitted to the runs it is validated against.

   Methodological

   The scalar correction V_CAL_K = 1.27 is the unweighted mean of per-run k_pre across V_50_01..04. Reporting per-run residual against these four runs is therefore a self-consistency check, not an independent validation. V_50_05 (run after baking V_CAL_K into firmware) is the first generalisation check. Multi-day / multi-board / multi-fluid generalisation is out of Rev-B scope.

6. §7

   Tethered 3.3 V supply.

   Scope

   The validation dataset is collected with 3.3 V supplied directly to MCU VDD via ST-Link, bypassing the TPS610981 boost converter from the single-cell battery design. The dataset validates the optical detection chain, sphere-model calibration, and dual-beam architecture. The field power chain (1 × AA Li-ion → boost → MCU) is deferred to Rev-C.

Per-drop scatter of TOP vs BOT pulse widths across the bench mount positions. The position-dependence finding's raw-waveform evidence is in the Analysis section above; this figure shows the derived 2D feature space the chord-time architecture's scalar V_CAL_K can't compress.

Alarm state machine.

Six states, two thresholds (±15% warning, ±25% alarm — per IEC 60601-2-24 §201.12). A nurse cannot disarm into Normal without explicitly Arming the device, preventing accidental silencing.

Assembly.

1. 01 PRINT

   Print the case

   Two FDM-printed shells in ASA white — front and back.

   Parts ASA white filament

   Tools FDM printer
2. 02 SOLDER ESD

   Solder the LCD + tactile switches

   EA DOGS164W-A LCD and two PTS647 tactile switches (left / right). Underlay a 3 mm spacer between the LCD and the PCB.

   Parts EA DOGS164W-APTS647 ×23 mm spacer

   Tools Soldering ironFlux
3. 03 SOLDER POLARITY

   Build the satellite breakout

   Press the two IR LEDs into their bezels first, then solder both bezeled LEDs to the satellite PCB. Solder the AA battery leads onto the same board.

   FIG · 03A IR LED polarity — both channels

   Parts VSLY5940 IR LED ×2IR bezel ×2AA battery + leadsSatellite PCB

   Tools Soldering ironHelping hands
4. 04 ASSEMBLY

   Attach the FFC cable

   Open the locking latch on J4 (6-pin FFC socket) on the main PCB, slide the ribbon in, close the latch. Repeat on the matching FFC socket on the satellite — no solder joint on either end.

   Parts 6-pin 0.5 mm FFC ribbon

   Tools Tweezers
5. 05 MECH

   Seat the PCB in the front case

6. 06 MECH

   Close the back case · insert the photodiodes

   Press each BPV10NF photodiode into its bezel, then push the bezeled photodiodes into the beam-aperture pockets in the back shell. Snap-fit the rear shell onto the front.

   Parts BPV10NF ×2IR bezel ×2Printed back shell

The choices behind the board.

Every non-trivial design choice is captured as a separate ADR in docs/decisions/. The cards below summarise the eight that matter most for the Rev-B submission. Click any card for the full record — context, alternatives considered, and the open risks at the time the decision was made.

What Rev-B proved.

Three architectural conclusions Rev-B forced. The next student doesn't relitigate these — each is bench-evidenced in Limitations and the long form lives in docs/limitations.md.

1. A single scalar V_CAL_K cannot generalise across mount positions.

   Architectural

   K required to fit gravimetric truth ranged 0.42–1.98 across 11 runs at four mount positions — same chamber, same drip set, same fluid, same board. Eight algorithm variants were trialled (TOP_low, BOT_low, BOT_core, TOP_core, mean-of-pulses, leading-half-pulse, hybrid splash-detector, EMA-adaptive); none passed ±5 % cross-position. Rev-C optics — not Rev-B firmware tweaks — are the path. See §17 in Limitations.

2. The mean-of-pulses architecture cannot recover drop oscillation phase.

   Fundamental

   Drops oscillate between oblate and prolate after detachment (Rayleigh 1879); the chord-time architecture catches each one at an unpredictable phase. The run-mean bias averages out (V_CAL_K = 1.27 works on the run mean) but the per-run residual stays at ±30 %. This is a physics floor of the single-vertical-chord topology, not a tuning problem — recovering phase needs horizontally-separated beam pairs (Rev-C). See §1 in Limitations.

3. Per-drop bench ground truth needs a closed end-orifice.

   Methodological

   The device measures drops at the drip-chamber orifice; the bench scale captures mass at the end-of-tubing orifice. Those two orifices have different surface-tension dynamics — the 2026-05-13 campaign saw end-orifice drop counts 1.5–2× the chamber-drop counts. Per-drop pairing against the scale is therefore invalid; the only valid bench ground truth is the run-mean V_true_chamber = mass / N_UART. Future bench rigs need a catheter-equivalent closed end-orifice for per-drop validation. See §2 in Limitations.